CN104899640A - Simulator and method for neural network - Google Patents

Simulator and method for neural network Download PDF

Info

Publication number
CN104899640A
CN104899640A CN201410606697.7A CN201410606697A CN104899640A CN 104899640 A CN104899640 A CN 104899640A CN 201410606697 A CN201410606697 A CN 201410606697A CN 104899640 A CN104899640 A CN 104899640A
Authority
CN
China
Prior art keywords
circuit
neuron
cynapse
signal
output terminal
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201410606697.7A
Other languages
Chinese (zh)
Other versions
CN104899640B (en
Inventor
徐志强
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Individual
Original Assignee
Individual
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Individual filed Critical Individual
Priority to CN201910669674.3A priority Critical patent/CN111291873A/en
Priority to CN201410606697.7A priority patent/CN104899640B/en
Priority to CN201910669575.5A priority patent/CN111275178A/en
Priority to CN201710012154.6A priority patent/CN107045649B/en
Priority to CN201910694308.3A priority patent/CN111353590A/en
Publication of CN104899640A publication Critical patent/CN104899640A/en
Application granted granted Critical
Publication of CN104899640B publication Critical patent/CN104899640B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/049Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/06Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/063Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means

Abstract

The present invention relates to a simulator and a simulation method for simulating the cerebral neural network in the neurobiology. According to the present invention, by researching and analyzing detail features of the working process of cerebral neurons and synapses and simulating and implementing the detail features through an electronic circuit, so that the working process of the neurons and the synapses can be more reasonably and more accurately simulated and the neural network with more excellent performance can be structured.

Description

The analogue means of neural network and method
Technical field the present invention relates to a kind of electronic technology, be specially analogy method and the analogue means of a kind of intrerneuron and cynapse thereof for simulating cerebral cortex, temporal lobe and peripheral region thereof relating to Neurobiology field, and the simulative neural network adopting these neuron analogue means and cynapse analogue means to form.
The working mechanism of background technology brain is one of most important scientific research.It is rely on the cynapse interlinked between neuron and neuron to realize transmission and the process of information that current research discloses brain, also a lot of architectural feature and the operational detail of single neuron and cynapse is disclosed to a certain extent, but how neuron and cynapse thereof realized on earth to the Premium Features of brain by annexation and signal transacting, be just still sciences problems.This is mainly because single neuron and cynapse can observe its structure and work by the dissection of in vitro neuron and microscope, but in vitro single or multiple neurons also cannot form signal transacting path, and we cannot observe the course of work of biological whole live body brain from microcosmic angle.So, even if there is research to disclose specific structural features and the work characteristics of neuron and cynapse, but any effect is play in whole neural network for them, how to work, how to be related with other neurons, especially how to form and realize cerebral function macroscopically, then often still can not get completely reasonably explaining again.
In order to the work of analog demenstration cranial nerve, and constitution realization has the artificial intelligence of memory and function and thinking thus, and increasing research at present adopts the method setting up mathematic simulated mode or neuron analogue means to imitate the work of cranial nerve.Especially in recent years, based on the discovery of the synaptic plasticity of the STDP characteristic of cynapse, and based on the proposition of the plastic learning and memory models of this cynapse STDP, a lot of applicant have submitted the patented claim of a lot of neuron and cynapse analogue technique.
The neuron analogue technique of current employing electronic circuit, is generally made up of input film integrating circuit (film potential for imictron is integrated), the threshold triggers circuit (threshold triggers for simulated action current potential) with threshold triggers, impulse output circuit (pulse for simulating the action potential that aixs cylinder exports exports) etc.The analogue technique of nerve synapse, generally be made up of presynaptic membrane input circuit, postsynaptic membrane output circuit and the mimic channel etc. with cynapse transmission characteristic, the characteristic that wherein cynapse transmission characteristic mimic channel can transmit according to chemical synapse, especially the plasticity meeting STDP characteristic of cynapse, adjusts output signal by input signal.The STDP characteristic of synaptic plasticity, Spike Timing-Dependent Plasticity, the i.e. synaptic plasticity of Temporal dependency.Current theory is thought: the STDP plasticity that cynapse is transmitted, also the transmission usefulness of i.e. cynapse, sequential with the spike of action potential neuronic after the presynaptic is correlated with, if the spike of presynaptic neuron is early than the spike of postsynaptic neuron, then produce synaptic trasmission (long-term potentiation, i.e. LTP) phenomenon, and the delay of spike less then cynapse transmission enhancement effect is larger, (for the ease of describing, this situation, we can call that its spike signal has LTP feature); If the spike of presynaptic neuron is slower than the spike of postsynaptic neuron, then produce cynapse and transmit long-term depression (long-term depression, i.e. LTD) phenomenon, and the delay of spike is less, it is larger that cynapse transmission weakens effect, (for the ease of describing, this situation, we can call that its spike signal has LTD feature).Obviously, the LTP characteristic sum LTD feature of spike signal, all belongs to the STDP plasticity that cynapse is transmitted.
The analogue technique of existing various neuron, cynapse and neural network thereof, Problems existing is:
1, current neuron analogue technique is focusing more in the simulation of sensory neuron, and few for the analogue technique of intrerneuron, and obviously, relates to the intrerneuron of memory and thinking, even more important for the working mechanism disclosing brain.
2, more note and imictron to the simulation of the integration of input signal, neuronic triggering and action potential, and for the different subtype of neuron action potential, particularly to brain why can break out different subtype action potential, form the substantive reason of different subtype action potential and the forming process of different subtype action potential, lack understanding and pay attention to, and applicant finds that these details are extremely important simulating when same neuron has thinking and memory two kinds of duties simultaneously.
3, in neuron analogue technique, the integration paid attention to " aixs cylinder-dendron " type excitatory synapse input signal more, and it is less to the processing procedure of the modularity synaptic input signal of " aixs cylinder-cell space " and " aixs cylinder-aixs cylinder " type, and this Complex Neural Network that structure hyperchannel multi information is intermodulated, particularly for the Premium Features of simulation brain, such as experience or mood (happy, frightened, the reward with bounties) impact on neuronal messages process is very important.
4, in cynapse analogue technique, do not fully take into account the cynapse of different transmission characteristic between brain zones of different and the neuron of difference in functionality, only adopt the cynapse of same characteristic to build simulative neural network, causing cannot precise arrangements and the neural network simulating a sophisticated functions.
5, especially existing cynapse analogue technique is substantially all the cynapse of simulating normal synaptic transmission usefulness of reaching maturity, have.And a lot of cynapses in fact existed in brain, before stimulating through related neural coordination of action, the just cynapse of a kind of " prefabricated property ", naturally grow in fetus and baby development process and generate, these " prefabricated property " cynapses not yet form effective cynapse and transmit usefulness, and learning and memory afterwards and thought process, the collaborative stimulation of neuronic excitation activity after the presynaptic, is just " activated " gradually to develop into and possesses effective cynapse and transmit usefulness.Existing cynapse analogue technique is not noted and is simulated this process.And applicant studies, and this process of discovery is remembered for explanation, long-term memory transforms to long-term memory, explain that the not same-action of neuron in declarative memory forming process of cortex and hippocampus is extremely important.
6, the existing neurn simulation network with memory function, adopt cynapse STDP plasticity analogue technique, the accurate sequential relationship of the spike pulse of presynaptic caudacoria is depended on due to STDP effect, in order to produce the spike signal meeting STDP sequential feature, prior art is in neurn simulation network, introduce clock control end (or claiming sequential control end), the multiple sequential control end of general needs in a neuron matrix unit, even if adopt clock bus to reduce clock control end, a neuron matrix unit also needs two to three clock control ends.Due to the matrix unit composition that neural network is generally enormous amount, so need a lot of clock control end, cause circuit and circuit complex distribution.And for the sequential control of the spike of clock control end, not only complexity but also need highly accurate, in the prior art, this can only adopt computer system or microprocessor, is realized, cause complex system by program.According to this kind of technology, the work of neurn simulation network, need to depend on a computer system to be controlled by manual program, its intelligent essence depends on program composition and formed, thus cannot really simulating nature formed human brain work, the level of intelligence achieved by it also must be subject to the restriction of its computer system and manual program.
Summary of the invention the object of the invention is openly such neuron analogy method and device, it can carry out work with two kinds of different working morphologies by imictron, thus realize memory and reflection two kinds of neuron activities, and make both realizations and switch without the need to relying on complicated sequential control.
The present invention discloses analogy method and the device of several cynapse simultaneously, they more can improve the work of several not homo-synapse in simulation brain, for realizing the nervous activity of difference in functionality in brain, a kind of analogy method and the device that can simulate cynapse imformation memory being converted into long-term memory from short-term memory are especially disclosed.
The present invention also discloses so a kind of simulative neural network, it adopts above-mentioned neuron analogue means and cynapse analogue means to set up, without the need to the control of computer system and manual program, the STDP plasticity that cynapse is transmitted just can be realized, thus the memory realized information and reflection process.Further the present invention discloses another kind of simulative neural network, can either realize the short-term memory to information and reflection process, can also realize the work short-term memory of information being converted into long-term memory.
The first neuron analogue means of the present invention, includes dendron input end (dendron for imictron inputs), signal processing module (part integrated signal for imictron and process) and aixs cylinder output terminal (aixs cylinder for imictron exports), as prior art, signal processing module includes film integrating circuit (for simulating the process integrated film input excitatory potential), film discharge circuit (after triggering for simulated action current potential film potential depolarization and form the action of refractory period, be also sometimes referred to as depolarization circuit), (the action of neuron triggering action potentials when the excitement for analogue membrane current potential is integrated and reached activation threshold value of threshold triggers circuit, in order to distinguish, first threshold trigger circuit are called in the present invention, its activation threshold value is V1), with action potential pulse-generating circuit (for the output pulse of simulated action current potential, in order to distinguish, the first action potential pulse-generating circuit is called) in the present invention, the general output circuit also including action potential, (there is no substantive analog functuion, just for isolating the action potential pulse signal amplifying action potential pulse-generating circuit and produce), its concrete annexation is: dendron input end is connected to the input end of film integrating circuit, the output terminal of film integrating circuit connects the input end of first threshold trigger circuit, and the output terminal of first threshold trigger circuit connects the input end of the first action potential pulse-generating circuit, the output terminal of the first action potential pulse-generating circuit is connected to aixs cylinder output terminal by output circuit, the output terminal of the first action potential pulse-generating circuit is connected to film integration discharge circuit simultaneously.
The invention is characterized in: described neuron analogue means also includes Second Threshold trigger circuit and the second action potential pulse-generating circuit, the input end of Second Threshold trigger circuit is connected to the output terminal of film integrating circuit, the output terminal of Second Threshold trigger circuit is connected to the second action potential pulse-generating circuit, the output terminal of the second action potential pulse-generating circuit connects aixs cylinder output terminal, and the output terminal of the second action potential pulse-generating circuit is also connected to dendron input end by a reverse channel transmission simultaneously.The activation threshold value V2 of Second Threshold trigger circuit setting is greater than the activation threshold value V1 of first threshold trigger circuit setting.
Second Threshold trigger circuit are used for the first producing cause and the biological significance thereof that imictron breaks out high activation threshold value action potential, how to work in reflection and memory two kinds of different conditions, be even more important for the same neuron of simulation.
As a kind of prioritization scheme, between the input end of first threshold trigger circuit and the output terminal of film integrating circuit, be added with a signal delay circuit.Make the output signal of film integrating circuit, be delivered to the time of the input end of first threshold trigger circuit, than the time being delivered to Second Threshold trigger circuit input end, have an of short duration delay.Signal delay circuit be arranged so that, when dendron input end inputs strong signal, Second Threshold trigger circuit can reliably trigger prior to first threshold trigger circuit.When the program is delivered to the AIS of aixs cylinder section far away for simulating from dendron input and the exciting signal integrated, need the time that one of short duration.This of short duration time delay in technology in the past, owing to not understanding its biological meaning, thus not pay attention to by people and simulate.And applicant notices, to the simulation of this time delay, the substantial connection between position and different subtype thereof that the integration transmission of exciting signal of dendron input and action potential break out can be simulated more accurately, more be consistent with the working condition of real cerebral neuron, simulate the operative scenario that cerebral neuron produces memory more accurately.
As another kind of prioritization scheme, the output terminal of the second action potential pulse-generating circuit is also connected to film discharge circuit simultaneously.The work of original film discharge circuit can have been come by the output of the first action potential pulse-generating circuit, but the output terminal of two pulse-generating circuits is controlled film discharge circuit separately, more can meet the course of work of neuron two kinds of action potentials, and make to there will not be two kinds of action potential pulses to repeat to produce the situation of superposition.
Further, described neuron analogue means also with one for switching " the attentional manipulation end " of memory/reflective operation state, (be with reflective condition or with remember condition carry out work for directly being controlled by external signal to switch neuron analogue means); Described Second Threshold trigger circuit, with threshold regulating circuit, for turning down the voltage triggered threshold value of Second Threshold trigger circuit, (making the trigger voltage V2 of Second Threshold trigger circuit be less than the trigger voltage V1 of first threshold trigger circuit); The input control end of threshold regulating circuit, is connected to " attentional manipulation end ".During work, " attentional manipulation end " can be connected to other neuronic aixs cylinder output terminals, modulated by other connected neuronic excitation activities, or, directly switch in high level and low level state, the two kinds of duties also namely directly making neuron analogue means process in imformation memory and reflection switch.When " attentional manipulation end " does not have input control signal, neuron is switched in " reflection " duty, carries out work according to foregoing working mechanism, and when working as " attentional manipulation end " input effective control signal of appearance, neuron is switched in " memory " duty, the voltage triggered threshold value being turned down Second Threshold trigger circuit by threshold regulating circuit makes it lower than the activation threshold value of first threshold trigger circuit, at this moment when the output signal of film integrating circuit is when constantly integrating rising, Second Threshold trigger circuit will prior to first threshold trigger circuit trigger action, make the second action potential pulse-generating circuit output action potential pulse, both aixs cylinder output terminal had been outputted to, also oppositely output to dendron input end simultaneously, form the spike signal required for cynapse STDP plasticity on dendron.This technical scheme is used for the second producing cause that accurate simulation neuron breaks out the action potential of high activation threshold value, described " attentional manipulation end " is to the adjustment of the setting voltage threshold value of Second Threshold trigger circuit, that simulation brain carries out " subconscious memory " (also i.e. intensive concentration) when working, by the neuronic activation of notice control path, the modulating action that (non-cynapse transmission) is transmitted in contact is created to the cell space of related Neurons or the cell membrane Na ion channel of AIS nearly cell space section, the action potential of high activation threshold value can be broken out at cell space or AIS nearly cell space section, produce action potential reverse transfer phenomenon, thus produce cynapse STDP plasticity, realize the memory function to information.Prior art is owing to lacking theoretic knowledge, so do not have associated analog technology to this Crack cause of outburst high threshold action potential.
As optimization, between " attentional manipulation end " and the input control end of threshold regulating circuit, also with the delay holding circuit of a duty.Postpone holding circuit to be used for making duty keep certain hour.This optimization, for simulating notice control path neuron, to the modulating action of information processing path neuron Na ion channel, has the maintenance effect of certain hour.
Accordingly, the first neuronic analogy method of the present invention, comprising: (1), to the signal of dendron input end input carry out integration; (integrating circuit namely corresponding to analogue means carries out integration to the input signal of dendron input end);
(2), detect the voltage signal after integration, (being equivalent to employing two threshold triggers circuit to detect the output terminal of integrating circuit);
(3) if this voltage signal is less than setting value 1, then neuron does not produce action, (the trigger voltage V1 that the output voltage being equivalent to integrating circuit is less than first threshold trigger circuit also i.e. setting value 1); If this voltage signal is equal to or greater than setting value 1 but its voltage rise slope is less than setting value 2, (rate of rise k=dv/dt, setting value 2 equals V2/T, wherein V2 is the trigger voltage of Second Threshold trigger circuit, the signal delay time of the signal delay circuit of T added by first threshold trigger circuit input end.The output voltage that this situation is equivalent to integrating circuit is greater than the trigger voltage V1 of first threshold trigger circuit, but the trigger voltage V2 of Second Threshold trigger circuit cannot be met or exceeded within the signal delay time that setting-up time T is also signal delay circuit), then trigger and produce action potential pulse, output to aixs cylinder output terminal, (being equivalent to first threshold trigger circuit triggers generation action potential and Second Threshold trigger circuit do not trigger), the voltage signal after integration is reset simultaneously, if this voltage signal is equal to or greater than setting value 1 and its voltage rise slope is equal to or greater than setting value 2, (output voltage being equivalent to integrating circuit is greater than the trigger voltage V1 of first threshold trigger circuit, and within setting-up time T, meet or exceed the trigger voltage V2 of Second Threshold trigger circuit), then trigger and produce action potential pulse, not only output to aixs cylinder output terminal but also oppositely output to dendron input end simultaneously, (be equivalent to Second Threshold trigger circuit triggers and produce action potential, not only be transported to aixs cylinder output terminal but also be transported to aixs cylinder input end by reverse channel transmission) simultaneously, voltage signal after integration is reset simultaneously.
Further, neuron analogy method of the present invention also switches memory and reflection two kinds of different operating states by the signal condition of " attentional manipulation end ", (be with reflective condition or with remember condition carry out work for being controlled by external control signal to switch neuron analogue means); When " attentional manipulation end " signal condition is " reflection ", neuron carries out work by aforementioned analog method; When " attentional manipulation end " signal condition is " memory ", then by above-mentioned the (3) step change into: if this voltage signal is less than setting value 1, then neuron does not produce action; If this voltage signal is equal to or greater than setting value 1, then trigger and produce action potential pulse, not only output to aixs cylinder output terminal but also oppositely output to dendron input end simultaneously, (be equivalent to " attentional manipulation end " turns down Second Threshold trigger circuit trigger voltage V2 by threshold regulating circuit, V2 is made to be equal to or slightly less than V1, at this moment when integral voltage reaches V1, Second Threshold trigger circuit have just triggered generation action potential, not only be transported to aixs cylinder output terminal but also be transported to aixs cylinder input end by reverse channel transmission simultaneously), the voltage signal after integration is reset simultaneously.
The second neuron analogue means of the present invention, includes dendron input end, signal processing module and aixs cylinder output terminal equally, as prior art, signal processing module includes film integrating circuit (for simulating the process integrated film input excitatory potential), film discharge circuit (after triggering for simulated action current potential film potential depolarization and form the action of refractory period), (the action of neuron triggering action potentials when the excitement for analogue membrane current potential is integrated and reached threshold value of threshold triggers circuit, in order to distinguish, be called first threshold trigger circuit in the present invention, its activation threshold value is V1), with the first action potential pulse-generating circuit (the output pulse for simulated action current potential), the general output circuit also including action potential, (for isolating the action potential pulse that amp pulse generation circuit produces), these all belong to prior art, its concrete annexation is: dendron input end is connected to the input end of film integrating circuit, the output terminal of film integrating circuit is connected to the input end of first threshold trigger circuit, and the output terminal of first threshold trigger circuit is connected to the input end of the first action potential pulse-generating circuit, the output terminal of the first action potential pulse-generating circuit connects aixs cylinder output terminal, the output terminal of the first action potential pulse-generating circuit is connected to film discharge circuit simultaneously, its technical characteristic is: described neuron analogue means is also with action potential reverse channel transmission, the output terminal of the first action potential pulse-generating circuit, is connected to dendron input end by this reverse channel transmission, the break-make of reverse channel transmission is subject to the control of a reverse transfer control circuit, the control input end of reverse transfer control circuit, is connected to one for switching " the attentional manipulation end " of neuron memory/reflective operation state.
The work of this neuron analogue means, adopt " attentional manipulation end " directly between imformation memory and reflection process two kinds of duties, to carry out switching to neuron analogue means to control, that is to say when " attentional manipulation end " input is effective, when being namely in " memory " state, by directly opening reverse transfer control circuit, make when neuron snapping motion current potential, action potential can carry out reverse transfer by reverse channel transmission to dendron end simultaneously, thus enable the synapse device be connected on dendron produce cynapse STDP plasticity, to realize the memory function to information.And work as " attentional manipulation end " when being in " reflection " state, close reverse transfer control circuit, action potential only exports to aixs cylinder and does not have reverse transfer, thus only produces the integration of input signal and reflection function and do not have memory function.
As optimization, between described " attentional manipulation end " and the control input end of reverse transfer control circuit, being also provided with the delay holding circuit of a duty, keeping certain hour for making duty.This is for simulating the neuron of notice control path, to the modulating action of the neuronic Na ion channel of information processing path, has the maintenance effect of certain hour.
Above-mentioned the second neuron analogue means, directly control neuronic action potential reverse channel transmission by " attentional manipulation end ", be the situation of the second generation action potential reverse transfer of imictron, namely brain realizes memory action by " subconscious memory " (being also intensive concentration).Certainly, on the basis of this, also can more simultaneously imictron another kind produce the situation of action potential reverse transfer, i.e. the exciting signal of dendron input strongly time, neuron is directly caused to break out the action potential of high activation threshold value, and to cell space and dendron reverse transfer.Realize the simulation of this working condition, neuron analogue means can be made to include Second Threshold trigger circuit again, the input end of Second Threshold trigger circuit is connected to the output terminal of film integrating circuit, and the output terminal of Second Threshold trigger circuit is connected to the control input end of reverse transfer control circuit.The high threshold action potential that Second Threshold trigger circuit break out for simulating cell space or the nearly cell space section of AIS, so the activation threshold value V2 of Second Threshold trigger circuit setting is greater than the activation threshold value V1 of first threshold trigger circuit setting.
Same, as the one optimization of such scheme, in this case, preferably between the input end and the output terminal of film integrating circuit of first threshold trigger circuit, be provided with a signal delay circuit.The program is delivered to the of short duration time delay of a section of the AIS section far away of aixs cylinder for simulating from the exciting signal of dendron input, and time delay is T.The technique effect of signal delay circuit follows the technology of the first neuron analogue means to be the same with meaning, has above and describes.
As another kind optimization, between the output terminal and the control input end of reverse transfer control circuit of Second Threshold trigger circuit, be also provided with a Time delay holding circuit.Its work is: when the output terminal of Second Threshold trigger circuit produces output signal, Time delay holding circuit is triggered and exports control signal and make output signal keep the certain time-delay time, thus makes reverse transfer control circuit open conducting and keep conducting state within a certain period of time by output signal.The pulse width of the action potential that Time delay time of Time delay holding circuit should produce slightly larger than action potential pulsing circuit, to make the reverse transfer process of action potential pulse, can be more complete more reliable.
The present invention discloses two kinds of neuron analogue means, its circuit structure and principle of work different.Wherein the first analogue means, action potential pulse to dendron reverse transfer produces circuit by Second Threshold to produce, and be that the activation threshold value turning down Second Threshold trigger circuit by threshold regulating circuit realizes directly circuit being switched in memory work state for switching memory or " the attentional manipulation end " of reflective operation state, this circuit structure and principle of work are closer to the working mechanism of cerebral neuron.And the second analogue means, the switch of control action current potential reverse channel transmission is directly carried out by " attentional manipulation end ", thus directly switch two kinds of duties of memory and reflection, the working machine of this circuit structure and principle of work and cerebral neuron is shaped with certain difference, but it follows the first analogue means to be the same to the result of input signal, its advantage is more suitable for realizing on the integrated, if particularly only need imictron to produce the second reason of memory function, i.e. " subconscious memory ", at this moment the control input end of reverse transfer control circuit only need be connected to external " attentional manipulation end ", then can without the need to arranging Second Threshold trigger circuit, be conducive to simplifying circuit.
Correspondingly, the second neuron analogy method of the present invention, comprising: (1), to the signal of dendron input end input carry out integration; (integrating circuit namely corresponding to analogue means carries out integration to the input signal of dendron input end);
(2), detect the voltage signal after integration; (being equivalent to the output end signal that first threshold trigger circuit detect integrating circuit), if this voltage signal is less than setting value 1, then do not produce action, (output voltage being equivalent to integrating circuit is less than the trigger voltage V1 of first threshold trigger circuit, and at this moment first threshold trigger circuit do not have trigger action);
If the voltage signal (3) after integration is greater than setting value 1, (output voltage being equivalent to integrating circuit is greater than the trigger voltage V1 of first threshold trigger circuit), then detect the signal condition of " attentional manipulation end "; At this moment,
If " attentional manipulation end " signal condition is " reflection " state, then neuron carries out work in " reflection " mode, neuron triggers and produces action potential pulse, output to aixs cylinder output terminal, (be equivalent to first threshold trigger circuit triggers, first action potential pulsing circuit produces action potential, and is sent to aixs cylinder output terminal), the voltage signal values after integration is reset simultaneously; (being equivalent to the action of film discharge circuit, the capacitor discharge by film integrating circuit);
If " attentional manipulation end " signal condition is " memory " state, then neuron carries out work in " memory " mode, (be equivalent to reverse transfer control circuit and open conducting), namely neuron triggers and produces action potential pulse, both output to aixs cylinder output terminal, oppositely output to dendron input end again; (be equivalent to first threshold trigger circuit triggers, first action potential pulse-generating circuit produces action potential, both be sent to aixs cylinder output terminal, be sent to dendron input end by reverse transfer control circuit again simultaneously), the voltage signal values after integration is reset simultaneously; (being equivalent to the action of film discharge circuit, the capacitor discharge by film integrating circuit).
Same conduct one is optimized, described " attentional manipulation end ", also with the delay holding circuit of a duty, for making duty keep certain hour, at this moment the signal condition detecting " attentional manipulation end " is changed into the state of the output terminal of testing lag holding circuit.This improvement, for simulating the neuron of notice control path, to the modulating action of the neuronic Na ion channel of information processing path, has the maintenance effect of certain hour.
Above-mentioned the second neuron analogy method, directly control to switch neuronic working method by " attentional manipulation end ", be the situation of the second generation action potential reverse transfer of imictron, namely brain realizes memory action by " subconscious memory " (being also intensive concentration).On the basis of this, also can more simultaneously imictron another kind produce the situation of action potential reverse transfer, namely the exciting signal of dendron input strongly time, directly cause neuron to break out the action potential of high activation threshold value, and to cell space and dendron reverse transfer.At this moment its analogy method also comprises further:
(4), detect the voltage signal after integration, if this voltage signal its voltage rise slope while being greater than setting value 1 is also greater than setting value 2, (voltage rise slope k=dv/dt, setting value 2 equals V2/T, wherein V2 is the trigger voltage of Second Threshold trigger circuit, the signal delay time of the signal delay circuit of T added by first threshold trigger circuit input end.The output voltage that this situation is equivalent to film integrating circuit is greater than the trigger voltage V1 of first threshold trigger circuit, and within time T, meet or exceed the trigger voltage V2 of Second Threshold trigger circuit), neuron is then made to carry out work in " memory " mode, (be equivalent to Second Threshold trigger circuit and export trigger pip, reverse transfer control circuit is made to open conducting), neuron triggers and produces action potential pulse, both output to aixs cylinder output terminal, oppositely output to dendron input end again; (be equivalent to the action potential of the first action potential pulse-generating circuit, be both sent to aixs cylinder output terminal, be sent to dendron input end by reverse transfer control circuit again simultaneously), the voltage signal values after integration is reset simultaneously; (being equivalent to film discharge circuit by the capacitor discharge of film integrating circuit).
And as the further improvement to two kinds of above-mentioned neuronic analogue means of the present invention or analogy method, described neuron analogue means is all also with modularity synaptic input circuit and modularity synaptic input end; Modularity synaptic input end is connected to the input end of modularity synaptic input circuit, and the output terminal of modularity synaptic input circuit is connected to the output terminal of film integrating circuit.Described modularity synaptic input circuit comprises two aspects: enhancement synaptic input circuit and enhancement synaptic input end, inhibitory synapse input circuit and inhibitory synapse input end thereof.This improvement makes neuron analogue means except having the dendron input end for excitatory signal input, also has the synaptic input function for carrying out modulating action to information processing.Notice that modularity input signal of the present invention is not connected on the input end of film integrating circuit, so it participate in directly unlike the excitatory signal of dendron input film potential integration and can direct triggering action potentials, but be connected on the output terminal of film integrating circuit, form a modulation voltage and enhancer or inhibitor is carried out to film integral output voltage, by the impact on film integral output voltage, can the exciting signal from dendron input and integration be modulated by triggering action potentials.This function is conducive to building the Complex Neural Network that hyperchannel multi information is intermodulated, to simulate experience or mood (happy, frightened, reward with bounties, the habituation etc.) impact on neuronal messages process.
Applicant is by researching and analysing two kinds of different subtypes of cerebral neuron snapping motion current potential, namely at the action potential of high activation threshold value of cell space or the outburst of AIS nearly cell space section, with the action potential of the low activation threshold value broken out in AIS section far away, the difference of both operational details, particularly research and analyse essence and the condition of both Crack cause, thus adopt and have more details and more reasonably neuron analogue technique, simulate the work of cerebral neuron more accurately, particularly when imictron switches the reflection of information and memory function, can be realized by simpler more reasonably mode.Above when describing concrete technical scheme, to technique effect and the positive effect of each concrete technical scheme, do corresponding explanation.
Coordinate the cynapse analogue means of neuron analogue means work, with reference to prior art, also can adopt the cynapse analogue technique of following discloses.Applicant researchs and analyses, and at least there is the cynapse of three kinds of different qualities in cerebral nerve network: " solidification cynapse ", " modifiable synapse " and " cynapse of prefabricated property ".The present invention adopts different technologies to simulate these cynapses respectively.
" solidification cynapse " of the present invention analogue means comprises cephacoria input end, caudacoria output terminal, it is characterized in that; Between cephacoria input end and caudacoria output terminal, be serially connected with a capacitor discharging circuit, capacitor discharging circuit simultaneously and be connected to an electric capacity leadage circuit.Formed with this and the most simply be there is the cynapse transmission channels that usefulness is transmitted in fixing cynapse, also namely do not possessed the cynapse transmission channels of synaptic plasticity.As the simplest scheme of one, capacitor discharging circuit is made up of a discharge capacity, and the front end of discharge capacity connects cephacoria input end, and the rear end of discharge capacity is transmitted resistance by one and connected caudacoria output terminal; Electric capacity leadage circuit is used for the electric charge of discharge capacity slowly to release; Transmitting reverse parallel connection on resistance and have a charging diode, for providing one oppositely to the particular path of discharge capacity reverse charging to the action potential pulse from postsynaptic membrane output terminal, transmitting the enhancing of usefulness for generation of cynapse.During work, if the input of cephacoria input end is the pulse of low frequency action potential, due to the effect of discharge capacity, after each action potential pulse is passed through, needing a period of time allows discharge capacity carry out electric discharge by leadage circuit could to recover.If continuous several low-frequency pulse, because discharge capacity is had little time electric discharge by charging, larger impedance can be presented to back to back pulse below, even if also cynapse transmission produces the phenomenon of cynapse suppression to low frequency action potential.If the input of cephacoria input end is high frequency spike pulse, because the impedance of electric capacity to high-frequency impulse is less, so impact is less.If particularly high frequency mo current potential can cause the outburst of the neuronal integration of postsynaptic membrane to have high threshold action potential, then because action potential meeting back transfer is to postsynaptic membrane, discharge capacity reverse charging is made by diode, thus make discharge capacity present the transmission usefulness obviously strengthened to next action potential pulse, even if also cynapse transmission produces cynapse enhancing (easily changing) phenomenon to high frequency mo current potential.
" solidification cynapse " of the present invention analogue means adopts extremely simple circuit to realize the cynapse propagation function of common cynapse, although strengthen just a kind of simply rough simulation to cynapse suppression and cynapse, but it is extremely simple to have circuit, thus the feature that cost and volume are extremely low is also had, can be applicable to the signal input and output path in simulative neural network, on the direct conduction paths of signal such as signal feedback path, obviously can save cost when simulative neural network larger.
In the prior art, adopt the modifiable synapse analogue means of electronic circuit, generally comprise cephacoria input end, caudacoria output terminal, STDP chronicle recognition circuit, STDP computing circuit, usefulness Circuit tuning is transmitted with cynapse, the spike signal that STDP chronicle recognition circuit can occur according to cephacoria input end and caudacoria output terminal, identify and judge whether its spike signal has STDP feature, and identify and judge the LTP feature that belongs to wherein or belong to LTD feature, the signal of these two kinds of features, be output to a STDP computing circuit and carry out plus-minus accumulation, output numerical value after computing transmits usefulness Circuit tuning for controlling cynapse, by transmitting usefulness Circuit tuning, the numerical value (that is to say the power that cynapse outputs signal) that usefulness is transmitted in cynapse is adjusted, change with the plasticity that this realizes transmitting cynapse output signal, namely synaptic plasticity is realized.Prior art adopts microprocessor to realize the above-mentioned function comprising STDP chronicle recognition circuit, STDP computing circuit and cynapse transmission usefulness Circuit tuning by program mostly.And the present invention discloses analogy method and the device of another kind of modifiable synapse.
The analogy method that the present invention has " modifiable synapse " of cynapse transmission STDP plasticity function comprises:
(1), controlled the output signal of caudacoria output terminal by the input signal of presynaptic membrane input end, make output signal synchronous with input signal; (2), detect the spike signal of presynaptic membrane input end and caudacoria output terminal; (3), when recognizing the spike signal occurring meeting STDP characteristic between cephacoria input end and caudacoria output terminal, then by a charge circuit and a discharge loop, discharge and recharge is carried out to an accumulate device according to the feature of signal: if signal is LTP feature, then charge, if signal is LDP feature, then discharge; The time width of charge or discharge is determined by STDP characteristic; (4), with the magnitude of voltage of accumulate device, adjust the size of voltage (or electric current) amplitude of cynapse output signal, and output to caudacoria output terminal; (5), accumulate device is set up the initialization voltage with certain numerical value when starting working.
As a kind of prioritization scheme, when the voltage on accumulate device is greater than or less than initialization voltage, slowly eliminate this voltage difference by a two-way leadage circuit, until the voltage on accumulate device equals initialization voltage value.(program simulation to the slow elimination of cynapse LTP and LTD effect, make cynapse analogy method of the present invention more perfect).
The analogue means that the present invention has " modifiable synapse " of cynapse transmission STDP plasticity function comprises cephacoria input end, caudacoria output terminal, STDP chronicle recognition circuit and transmits usefulness Circuit tuning.STDP chronicle recognition circuit belongs to prior art, the spike signal that can occur according to cephacoria input end and caudacoria output terminal, identify and judge whether its spike signal has STDP feature, and identification is judged belong to LTP feature or belong to LTD feature, in addition, STDP chronicle recognition circuit all has magnitude of voltage to the signal that cephacoria input end and caudacoria output terminal input and detects, magnitude of voltage due to the spike of action potential is obviously greater than the signal voltage value after cynapse is transmitted, so a STDP chronicle recognition circuit action potential spike higher to voltage amplitude identifies, and for the cynapse output signal that cynapse self produces, and the cynapse output signal of other synapse device to pass over from caudacoria output terminal, can not be identified because signal voltage amplitude is lower.STDP chronicle recognition circuit can adopt the microprocessors such as CPU or MCU, carries out analog to digital conversion to the spike signal of cephacoria input end and caudacoria output terminal, then according to its sequential priority and mistiming, identifies and export LTP characteristic signal and LTD characteristic signal.Also voltage comparator and time base circuit be can adopt, according to the sequential of spike signal successively and the mistiming, LTP characteristic signal and LTD characteristic signal exported by logic identification.This belongs to prior art, has cynapse and transmits in the cynapse analogue technique that STDP plastic employing electronic circuit realizes disclosed and have correlation technique.And, the work characteristics about STDP chronicle recognition circuit described here and actualizing technology, all equally applicable in other parts of the present invention.
The feature of " modifiable synapse " of the present invention is: STDP chronicle recognition circuit is with LTP output terminal and LTD output terminal; (there is the priority sequential control LTP output terminal of spike and the output signal of LTD output terminal according to cephacoria input end and caudacoria output terminal in described STDP chronicle recognition circuit); LTP output terminal and LTD output terminal are connected to an accumulate device for store electrical energy respectively by a charge circuit and a discharge loop.STDP chronicle recognition circuit LTP output terminal when identifying LTP feature occurs that LTP pulse exports, and the time interval of the pulse width of LTP pulse and former and later two spike signals presents negative correlation, also namely meet STDP characteristic, LTP pulse is charged to accumulate device by charge circuit; STDP chronicle recognition circuit LTD output terminal when identifying LTD feature occurs that LTD pulse exports, and the time interval of the pulse width of LTD pulse and former and later two spike signals presents negative correlation, and LTD pulse is discharged to accumulate device by discharge loop.
Accumulate device is connected with an initialization charging circuit; (initialization charging circuit is used for making accumulate device be charged to initialization voltage rapidly when circuit is started working, possesses initial quantity of electricity, cynapse is made just to possess transmission usefulness in original state, its initialization voltage that is to say the reference voltage of accumulate device, and usefulness is transmitted in the cynapse that at this moment cynapse exports as standard).
The output of accumulate device is connected to cynapse and transmits usefulness Circuit tuning, as the reference voltage that it exports; The output signal that usefulness Circuit tuning is transmitted in cynapse is connected to caudacoria output terminal as the output of cynapse signal.The synchro control that usefulness Circuit tuning accepts cephacoria input end input signal is simultaneously transmitted in cynapse, and its output signal by the modulation of accumulate device output voltage, is subject to the synchro control of cephacoria input end input signal in time in intensity.The output terminal of cynapse transmission usefulness Circuit tuning can also include the circuit of above-mentioned " solidification cynapse " analogue means, makes it the transmission characteristic simultaneously with cynapse enhancing and cynapse suppression.
The analogy method of " modifiable synapse " of the present invention and device, simulate the transmission characteristic of the chemical synapse with synaptic plasticity, but it adopts and changes the method that circuit transfer impedance (variable resistance circuit) realizes transmitting usefulness unlike prior art, but identify LTP and LTD feature by STDP chronicle recognition circuit, and by charging circuit and discharge circuit, discharge and recharge is carried out to an accumulate device, the voltage on accumulate device is made to produce height change along with charging (during LTP) and electric discharge (during LTD), and for modulating the output of synaptic efficacy Circuit tuning, thus the magnitude of voltage making cynapse export pulse produces the corresponding adjustment of height, the simulation to STDP synaptic plasticity is realized with this.Cynapse analogue means of the present invention, it is also the adjustment of cynapse transmission usefulness to synaptic plasticity, what embody is output voltage change instead of output impedance change, so be more convenient for various neuron analogue means cooperating, be particularly suitable for making autonomous device, the simulation demonstrating apparatus of constructing neural network, and the current transmission usefulness of this cynapse can be shown in real time conveniently by the voltage value detecting accumulate device.And, do not need the variable resistance circuit of programming, closer to the real work situation of " natural ".The analogue means of " modifiable synapse " of the present invention, simulation be the nerve synapse being present in the positions such as the hippocampus of brain, corpus straitum, amygdaloid nucleus and partial thickness region, having the very strong and STDP synaptic plasticity that can produce fast, is the key character forming journey and long-term memory in short-term.
As one optimization, accumulate device is also connected with a two-way leadage circuit; Two-way leadage circuit is used for slowly eliminating accumulate device causes voltage rising or reduction change due to discharge and recharge within a certain period of time, until the voltage on accumulate device equals initialization voltage.Thus the slow elimination of simulating cynapse LTP and LTD effect, make cynapse analogue means of the present invention more perfect, the analog operation of continuous multi-process can be adapted to.As the simple scheme of one, two-way leadage circuit comprises a mu balanced circuit, and (voltage of mu balanced circuit equals the initialization voltage of accumulate device, is also reference voltage), mu balanced circuit is connected to accumulate device by a resistance.When the voltage of accumulate device is greater than mu balanced circuit setting voltage, (when also namely there is cynapse LTP effect), accumulate device is slowly discharged to mu balanced circuit by resistance, (also namely slowly eliminating LTP effect); When the voltage of accumulate device is less than voltage of voltage regulation setting voltage, (when also namely there is cynapse LTD effect), mu balanced circuit is slowly charged to accumulate device by resistance, (also namely slowly eliminating LTP effect).After a certain time, after the term of validity of also i.e. synaptic plasticity, the voltage of accumulate device is made to equal mu balanced circuit setting voltage, normal output amplitude when namely plasticity does not appear in cynapse yet.
The analogy method of " cynapse of prefabricated property " of the present invention comprising: (1), between the cephacoria input end and caudacoria output terminal of simulation cynapse preset a cynapse transmission channels, but the work of this cynapse transmission channels of closedown that begins of work; (2), detect the spike signal of presynaptic membrane input end and caudacoria output terminal; (3), between cephacoria input end and caudacoria output terminal, there is the spike signal meeting STDP characteristic, then computing is carried out to the appearance situation of this signal, (comprising integration or counting); (4), when the satisfied setting rule of situation that the spike signal meeting STDP characteristic occurs, then start the work of default cynapse transmission channels, the cephacoria input end of cynapse and caudacoria output terminal are set up there is the transmission channels that usefulness is transmitted in cynapse.
Described setting rule is the one in following two kinds: 1, only counting meets the number of times of the signal of the LTP feature in STDP characteristic, if within the time of one section of setting, occurs that the counting of LTP reaches setting value, then starts the work of default cynapse transmission channels; Or 2, do not set gate time, counting meets the number of times of the signal of LTP characteristic sum LTD feature simultaneously, wherein occurs that the signal number of times of LTP feature does addition, occurs that the signal number of times of LTD feature does subtraction; If tale reaches setting value, then start the work of default cynapse transmission channels.
Further improvement, the analogy method of " cynapse of prefabricated property " of the present invention, also comprises:
(5), after the work starting the cynapse transmission channels preset, continue the spike signal detecting presynaptic membrane input end and caudacoria output terminal; (6), between cephacoria input end and caudacoria output terminal, there is the spike signal meeting STDP characteristic, then computing is carried out to the appearance situation of this signal; When the spike signal meeting STDP characteristic occurs that situation meets the second setting rule, then close the work of default cynapse transmission channels, cut off the cephacoria input end of cynapse and the transmission channels of caudacoria output terminal.
The second described setting rule is the one in following two kinds: 1, only counting meets the number of times of the signal of LTD feature, if within the time of one section of setting, occurs that the counting of LTD reaches setting value, then closes the work of default cynapse transmission channels; 2, counting meets the number of times of the signal of LTP characteristic sum LTD feature simultaneously, wherein occurs that the signal number of times of LTP feature does addition, occurs that the signal number of times of LTD feature does subtraction; If tale is lower than setting value, then close the work of cynapse transmission channels.Obviously in rear a kind of this setting rule, close the setting value of the tale of cynapse transmission channels, the setting value of the tale starting cynapse transmission channels must be less than.
The analogue means of " cynapse of prefabricated property " of the present invention, the cynapse transmission channels including cephacoria input end, caudacoria output terminal and preset; It is characterized in that: also with a STDP chronicle recognition circuit, for identifying the spike signal between cephacoria input end and caudacoria input end with STDP feature; The output of STDP chronicle recognition circuit is connected to a STDP integrating circuit, and the output of STDP integrating circuit is connected to a thresholding control circuit; The output terminal of thresholding control circuit is connected to a channel switching circuit; Channel switching circuit is for controlling the break-make of cynapse transmission channels.
Its course of work is: STDP integrating circuit, according to setting rule, is accumulated the signal that the institute of STDP chronicle recognition circuit within a certain period of time identifies output, and is converted into its output valve outputed signal; Thresholding control circuit, for detecting the output valve of the output signal of integrating circuit, reaches when the output valve of integrating circuit or higher than setting thresholding 1, then produces and export control signal; And start connection cynapse transmission channels by channel switching circuit, set up between cephacoria input end and caudacoria output terminal and there is the cynapse transmission channels that usefulness is transmitted in cynapse.
Its course of work further comprises: after channel switching circuit starts connection cynapse transmission channels, thresholding control circuit continues the output valve of the output signal detecting integrating circuit, when the output valve of integrating circuit is lower than setting thresholding 2, then closes and export control signal; And close cynapse transmission channels by channel switching circuit.Also the cynapse transmission channels between cephacoria input end and caudacoria output terminal is namely cut off.Obviously, set thresholding 2 to need to be less than setting thresholding 1.More reasonably arrange as one, setting thresholding 2 can be equivalent to 65% to 85% of the value setting thresholding 1.
Described default cynapse transmission channels, be one there is cynapse to transmit the cynapse transmission channels of usefulness, one or more in the synaptic plasticity that these cynapse transmission characteristics comprise synaptic facilitation, cynapse suppression, cynapse Long-term plasticity particularly have STDP characteristic, these transmission characteristics.The cynapse transmission channels preset is exactly in fact a cynapse analogue means with transmission characteristic, it can be aforesaid " fixing cynapse " device, also can be aforesaid " modifiable synapse " device, or the cynapse analogue means of other prior aries, comprise the Synaptic junction device adopting memory varistor material to make.
The setting rule of described integrating circuit is: the output valve improving output signal when there is the signal of LTP feature, reduces the output valve of output signal when there is the signal of LTD feature.(relative to analogy method, analogue means only adopts wherein a kind of mode closer to brain work characteristics to the operation rule of STDP).
The analogy method of " cynapse of prefabricated property " of the present invention and device, transmit compared with the plastic cynapse analogue technique of STDP with the cynapse that has of prior art, although all have the technology meeting STDP characteristic spike signal between detection and Identification presynaptic caudacoria each other, its process made is different.Existing STDP cynapse analogue technique, initially there is cynapse transmit usefulness, again according to detecting the spike signal meeting STDP characteristic identified, adjust cynapse and transmit usefulness: if there is LTP characteristic signal, then promote cynapse and transmit usefulness, if there is LTD characteristic signal, then reduce cynapse and transmit usefulness, realize cynapse with this and transmit plasticity.From brain structure, this class cynapse is present between the neuron in the regions such as hippocampus, amygdaloid nucleus, corpus straitum more, for realizing the short-term memory (or claiming long-term memory) of information.
And " cynapse of prefabricated property " of the present invention, itself just include one and possess the cynapse transmission channels that usefulness is transmitted in cynapse, also namely a cynapse analogue means itself is included, be on the basis of cynapse analogue means, add the technology how identifying and judge signal, carry out starting or closing cynapse transmission channels again with this.Its cynapse transmission channels when working initial is closed, do not possess cynapse and transmit usefulness, only has the spike signal repeatedly occurring the LTP feature met in STDP characteristic between cephacoria input end and caudacoria output terminal, and accumulation reaches a certain degree, just can start the work of connecting cynapse transmission channels, between cephacoria input end and caudacoria output terminal, set up effective cynapse transmit usefulness.What its was simulated is a kind of cynapse be present between the neuron on the regions such as cortex, do not possess cynapse when this kind of cynapse starts to be formed and transmit usefulness, need after the neuronic collaborative stimulation repeatedly for a long time in front and back, just be activated into effective cynapse, and once activate, it is just more stable that it transmits usefulness.
" cynapse of prefabricated property " is once be activated become effective cynapse, it just can not be closed automatically, if also there is the spike signal of LTP feature, as long as or have and to be made by its cynapse transmission presynaptic neuron be activated to produce action potential (actual this is also produce LTP feature), then the numerical value of STDP integrating circuit is just more and more higher until reach maximal value, make cynapse transmission effects more effectively reliable, and even without occurring LTP signal, cynapse also can continue the transmission usefulness of remaining valid.Only have and do not occur LTP feature for a long time, and repeatedly occur the inhibition signal of LTD feature, the numerical value of integrating circuit just can be made slowly to reduce, final lower than threshold value, thresholding control circuit cuts out output signal, closes cynapse transmission channels, cynapse was lost efficacy again by channel switching circuit.This duty is for simulating the stability of brain long-term memory, even but information is preserved as long-term memory, if these information do not use for a long time, and have obscuring for a long time (competition) of other inhibition information, then long-term memory also can slowly pass into silence.
" cynapse of prefabricated property " of the present invention analogue means, STDP chronicle recognition circuit belongs to prior art, and how to LTP and the LTD characteristic signal that STDP chronicle recognition circuit exports, computing is carried out for STDP integrating circuit, generally the microprocessors such as MCU can be adopted, according to setting rule, count operation is carried out to LTP characteristic signal and LTD characteristic signal, then count value is compared with setting value, export control signal according to comparative result.The present invention is the open another kind of analogue technique adopting accumulate device to realize STDP integrating circuit also.Its STDP integrating circuit includes charging circuit, discharge circuit and an accumulate device, and the LTP output terminal of STDP chronicle recognition circuit and LTD output terminal are connected on accumulate device respectively by charging circuit and discharge circuit; The output terminal of accumulate device is connected to the input end of thresholding control circuit.During work, there is adjustment in the voltage of accumulate device, thresholding control circuit detects its voltage, compares with setting value, and opens or close the work of cynapse transmission channels with this along with STDP effect.Can realize equally being activated by repeatedly LTP characteristic signal starting cynapse transmission channels, " cynapse of prefabricated property " is become " effective cynapse "; Also can close cynapse transmission channels by LTD characteristic signal repeatedly again, " effective cynapse " was lost efficacy, the extinction of simulation cynapse.This mode without the need to adopting MCU, also without the need to relying on manual program to realize STDP integral operation, more intuitively can more naturally simulate the course of work of human brain.
The analogue means of " cynapse of prefabricated property " of the present invention, be particularly suitable for being made into independently cynapse analogue means, with neuron analogue means disclosed by the invention, form experimental neurn simulation network, for demonstrating and study work and the function of brain, and neuron analogue means can freely be connected as required with cynapse analogue means, without the need to the restriction by topological structure, neurn simulation network can be formed by any topological structure.When this application, the analogue means of described " cynapse of prefabricated property ", also can with a voltage display device, for showing the output voltage of accumulate device.Thus intuitively can show the degree of strength of the transmission usefulness of this cynapse, be more conducive to scientific experiment.
Adopt above-mentioned neuron and cynapse analogue means, the simulative neural network with different working characteristic can be set up.
Such as, a kind of neurn simulation network with imformation memory and reflection function, comprise at least one neuron matrix unit, each neuron matrix unit comprises multiple (multiple i.e. two or more) presynaptic neuron analogue means and multiple postsynaptic neuron analogue means, between the aixs cylinder output terminal of presynaptic neuron analogue means and the dendron input end of postsynaptic neuron analogue means, employing has the plastic cynapse analogue means of cynapse transmission usefulness STDP and connects, it is characterized in that: " the attentional manipulation end " of each neuron analogue means links together, by a road, there is low and high level two states, control signal for the memory/reflection two kinds of duties switching neuron analogue means controls.
Neurn simulation network of the present invention, can by neuron analogue means and between connect there is the plastic cynapse analogue means of cynapse STDP, realize the memory to information and reflection.And, compared with existing neurn simulation network, neurn simulation network of the present invention, no matter be made up of how many neurons, also how huge the neuron matrix unit quantity no matter formed have, all neuronic " attentional manipulation end " need link together, be controlled by the control signal of same road memory/reflective operation state by all the same.When control signal is placed in " reflection " state, neurn simulation network carries out reflection process (also namely simulating thinking and the reaction of brain) in the mode of reflection to input information, only to the unidirectional output of aixs cylinder during neuron snapping motion current potential, cynapse analogue means can not produce transmission usefulness STDP plasticity owing to lacking the spike signal of caudacoria, and the processing procedure of information can not cause interference to the recall info (experience, rule, custom etc.) originally formed; When control signal is placed in " memory " state, the neuron of neurn simulation network processes (learning and memory of also namely simulating brain) input information with the working method of memory, both to axonal transport during neuron snapping motion current potential, also to dendron back transfer, cynapse analogue means is enable to produce STDP plasticity.In other words, neurn simulation network of the present invention is when carrying out imformation memory, the pulse of spike accurately required for its generation cynapse STDP plasticity and sequential thereof, (as the brain) that automatically produced by that neuron activated by excitement oneself, and without the need to producing additionally by other circuit.So, neurn simulation network of the present invention is without the need to arranging many clock control ends, more accurately control without the need to the sequential of manual program to the spike of clock control end of being unified by department of computer science, so, the system architecture of neurn simulation network of the present invention is simpler, also carrying out work without the need to depending on department of computer science's manual program of unifying, the work of the brain of natural evolution can be simulated more veritably.
The present invention also openly another kind there is the neurn simulation network of two kinds of different time courses memories, it has short-term memory and long-term memory, and short-term memory can be converted into long-term memory in any case.Described simulative neural network comprises the first input neuron group, the first output neuron group, the second input neuron group, the second output neuron group, the first intrerneuron group and the second intrerneuron group;
The aixs cylinder output terminal of the first input neuron group is connected with adopting " solidification cynapse " between the dendron input end of the first output neuron group, and the aixs cylinder output terminal of the second input neuron group is connected with adopting " solidification cynapse " between the dendron input end of the second output neuron group; (neuron of the stable information input and output passage formed in simulation brain connects);
The aixs cylinder output terminal of the first input neuron group and the second input neuron group, and between the dendron input end of the first intrerneuron group, adopt " modifiable synapse " to connect; (simulation cortex or the neuron of smelling cortex by the projection to the dentate granule cell of hippocampus of the path of perforating fiber, and adopt between the two there is the plastic cynapse of STDP connect);
The aixs cylinder output terminal of the first intrerneuron group, and between the dendron input end of the second intrerneuron group, adopt " modifiable synapse " to connect; (simulate inside hippocampus, the mossy fiber that granular cell sends, and the connection with the plastic cynapse of STDP between hippocampus CA3 cones);
The aixs cylinder output terminal of the second intrerneuron group, and between the dendron input end of the first output neuron group and the second output neuron group, adopt " modifiable synapse " to connect; (Schaffer that sends of CA3 cones of simulation hippocampus is collateral, and has the plastic Synaptic junction of STDP between CA1 cones);
Between the aixs cylinder output terminal of the first input neuron group and the dendron input end of the second output neuron group, between the aixs cylinder output terminal also having the second input neuron group and the dendron input end of the first output neuron group, " cynapse of prefabricated property " is adopted to connect; (simulation is present in the connection of the prefabricated property cynapse between brain some neuron particularly on cortex, and can activate in any case and become effective cynapse, thus simulated implementation short-term memory is to the conversion of long-term memory).
Simulative neural network of the present invention, can realize between two-way information IO channel, short-term memory is formed fast by " modifiable synapse " of intrerneuron and connection thereof, (also long-term memory can be claimed, or claim hippocampus memory, namely by memory that people's cerebral hippocampal realizes), and, if this short-term memory is repeatedly started and is remembered, then can activate and be preset at " cynapse of prefabricated property " between two information IO channel, thus between two information IO channel, set up more directly more stable cynapse transmission channels, realize long-term memory.(also claim permanent memory, or claim cortex memory, what be namely recorded in the regions such as cerebral cortex can the memory of longer-term more stable existence).And once set up long-term memory (cortex memory) by " cynapse of prefabricated property ", then original synaptic channel passing through the short-term memory (hippocampus memory) that " modifiable synapse " is set up, then can due to the foundation of " cynapse of prefabricated property " this more direct transmission channels, and make " modifiable synapse " this indirect path always occur that cynapse suppresses, final forfeiture effect, and discharge the short-term memory for re-establishing new information.
Simulative neural network of the present invention, if as a kind of experiment and apparatus for demonstrating, the process that the formation of neuronic reflex activity, short-term memory in brain and short-term memory are converted into long-term memory can be simulated better, be conducive to the working mechanism of the learning and memory disclosing brain better.Its working mechanism, conversion process and biological significance thereof separately have later describe.
Simulative neural network of the present invention, if be applied to some automation field with circuit mode, monoenergetic is enough between two groups of signals, set up memory relation (short-term memory) quickly according to its correlativity (sequential relationship of signal) for it, thus by a kind of signal, the automatic control of adaptivity is carried out (as working rule to another kind of signal according to memory, working routine etc.), namely also there is study and adaptive ability.And, if working time long enough, its automatic control function can also slowly form and show certain " custom " according to the incidence relation repeatedly occurred between signal, (as features such as experience, custom and personality that human brain is formed due to Term Learning), has unique adaptive control effect.
Description of drawings 1 is neuronic structural representation.Fig. 2 is the circuit block diagram of the neuron analogue means of prior art.Fig. 3 is the principle of work schematic diagram of the neurn simulation network of prior art.Fig. 4 is the structural representation of neuronic aixs cylinder.Fig. 5 is the circuit block diagram of the first neuron analogue means of the present invention.Fig. 6 adopts single-chip microcomputer to realize the circuit theory diagrams of the neuron analogue means of Fig. 5.Fig. 7 adopts common simulation circuit to realize the circuit theory diagrams of the neuron analogue means of Fig. 5.Fig. 8 is the circuit block diagram of the second neuron analogue means of the present invention.Fig. 9 adopts single-chip microcomputer to realize the circuit theory diagrams of the neuron analogue means of Fig. 8.Figure 10 adopts common simulation circuit to realize the circuit theory diagrams of the neuron analogue means of Fig. 8.Figure 11 is the circuit theory diagrams of the modularity input circuit of neuron analogue means.Figure 12 is the circuit theory diagrams of " solidification cynapse " of the present invention.Figure 13 has the schematic diagram that usefulness and spike sequential relationship are transmitted in the plastic cynapse of STDP.Figure 14 is the circuit block diagram with the plastic cynapse analogue means of STDP of prior art.Figure 15 is the circuit block diagram of " modifiable synapse " of the present invention analogue means.Figure 16 is the circuit theory diagrams of a kind of embodiment of Figure 15 " modifiable synapse " analogue means.Figure 17 is the circuit block diagram of " cynapse of prefabricated property " of the present invention analogue means.Figure 18 is that a kind of accumulate device that adopts is to realize the circuit block diagram of STDP integrating circuit " cynapse of prefabricated property " analogue means.Figure 19 is the circuit theory diagrams of Figure 18 " cynapse of prefabricated property " analogue means.Figure 20 is the circuit theory diagrams of a kind of embodiment of " cynapse of prefabricated property " analogue means.Figure 21 is a kind of circuit theory diagrams of STDP identification circuit.Figure 22 is the principle of work schematic diagram of a kind of neurn simulation network of the present invention.Figure 23 is the principle of work schematic diagram of another kind of neurn simulation network.
Be to principle of the present invention and describing of specifically implementing below embodiment.
According to the research of current Neurobiology, typical intrerneuron also namely the composition of neurocyte as Fig. 1, form primarily of cell space 1, dendron 2 and aixs cylinder 3.Neuron generally has many dendrons, and dendron is branch repeatedly, and shape is as branch; Cell space and dendron all can form cynapse with from other neuronic Nerve Terminals In The Human Skins above, become neuronic input end, receive from neuronic signal above.Intrerneuron generally has an aixs cylinder 3, and the tip of aixs cylinder 3 forms cynapse with other neuronic dendrons below or cell space again, and aixs cylinder 3 is equivalent to neuronic output terminal, neuron transmission of signal rearwards.
Cynapse (mainly referring to chemical synapse here) is the important step of signal transmission between neuron.Cynapse comprises presynaptic membrane, postsynaptic membrane and synaptic cleft, and what be in last neuronic Nerve Terminals In The Human Skin is presynaptic membrane, and after being in, a neuronic dendron or cell space is postsynaptic membrane.Current neural unit be excited life birth raw action potential time, presynaptic membrane is to synaptic cleft release neurotransmitters, and postsynaptic membrane absorbs neurotransmitter and causes the change of inside neurons ionic species, thus produces the electrical excitation of cell membrane.These electrical excitation are carried out superposition in inside neurons and are integrated, comprise the spatial integration of the transmission of signal from multiple cynapse, integrate with the time of the multiple signals from same cynapse, when the electrical excitation integrated meets or exceeds a certain threshold value, neuron snapping motion current potential.Action potential is delivered to Nerve Terminals In The Human Skin, passes to next neuron by cynapse, settling signal integration, triggering and transmittance process, eliminates, the residual excitement of film to carry out signal integration next time simultaneously.
The signal transmission usefulness of cynapse can change, and namely cynapse has plasticity, and this is the basis of learning and memory.According to current widely accepted theory, the plasticity change that usefulness is transmitted in cynapse follows STDP principle, i.e. the synaptic plasticity of Temporal dependency, specifically sees the description of the present invention's " background technology " part.
Fig. 2 is that the circuit of at present typical intrerneuron analogue technique is formed, generally by input end (stimulation of simulation dendron or cell space inputs), signal processing module (part integrated signal for imictron and process) and output terminal (exporting for simulating aixs cylinder), signal processing module wherein includes input film integrating circuit (process for analogue membrane excitatory potential is integrated), the threshold triggers circuit action of neuron triggering action potentials (when the integration for analogue membrane excitatory potential reaches threshold value), pulse-generating circuit (pulses generation for simulated action current potential), with film discharge circuit (trigger the depolarization of caudacoria excitatory potential for simulated action current potential and form the action of refractory period), general, also include the output circuit of action potential, for isolating the action potential pulse that amp pulse generation circuit produces.This is the core composition of current neuron analogue technique, various neuron analogue technique disclosed at present, although circuit composition has various ways, to the signal transacting of input end, to the waveform modelling of action potential, to the depolarization of film excitement and the accurate analog of refractory period, etc., there is a lot of relevant improvement opportunity, but the core of its neuron simulation, still by above-mentioned structure is formed.
Fig. 3 typically adopts neuron mimic channel to form the isoboles of the neuron matrix of neural network at present.In a neurn simulation network, be made up of the neuron matrix unit of n × n n the input neuron analogue means (A1 to An) being in column position and n output neuron analogue means (C1 to Cn) being in line position, some is also with n input neuron analogue means (B1 to Bn) of another input channel.Be formed between neuron axon and other neuronic dendrons and simulate cynapse PY, simulation cynapse generally adopts variable resistance technology, has STDP effect.The acting in conjunction of the accurate sequential of the spike of presynaptic caudacoria is depended on due to cynapse STDP effect, the spike of presynaptic membrane that is to say the signal of the action potential of the aixs cylinder of presynaptic neuron, and in order to produce the spike signal of postsynaptic membrane, prior art is generally in neuron analog network, introduce clock control end (or claiming sequential control end), as CLK-A, CLK-B and CLK-C in Fig. 3, produce the postsynaptic membrane spike signal meeting STDP and require according to the needs of signal sequence.Neuron matrix can be made to realize STDP effect thus realize memory function although do like this, but Problems existing is: 1, need to introduce clock control end (or claiming sequential control end), n sequential control end is generally needed in a neuron matrix unit, even if adopt clock bus to reduce clock control end, a neuron matrix unit also needs 2 to 3 clock control ends.Due to neuroid be generally enormous amount matrix unit composition, so need that there is a lot of clock control end, cause circuit structure and circuit distribution complicated.2, because cynapse STDP plasticity is for the high susceptibility of the time sequential of front and back film spike, slightly mistake just can affect the effect of synaptic plasticity, even produce a contrary effect, so, for the sequential control of the spike that clock control end exports, need with the work schedule of each neuron analogue means closely related, not only complexity but also need highly accurate, in the prior art, this can only adopt computer system or microprocessor CPU, realized by the computer program of manual compiling, cause complex system.3, the analogue technique of neural network, originally wish simulated by analog network and realize artificial intelligence, and conventionally, the work of the analog network of this artificial intelligence, also need to depend on a computer system to be controlled by the manual program of complexity, obviously, this cannot the neurn simulation network of divorced from computer system and manual program, the human brain work that really cannot simulate self-assembling formation, more can hinder the raising of the level of intelligence of simulation artificial intelligence, (being subject to the restriction of the design level of computer system and manual program thereof).This is also existing neuron and the problem existing for neuron network simulation technology.
Applicant is by studying further the formation of neuron action potential and generation mechanism, notice that neuron produces in the process of action potential at integration input stimulus (film is excited), can due to the trickle difference of the change shape of input stimulus signal, and the chemical substance environment of neuron outside is by the impact of other neuron action of surrounding, and form the action potential producing different subtype with different producing methods, and bring the gross differences of working result.According to these results of study, applicant devises more perfect neuron analogue technique, solves above-mentioned the deficiencies in the prior art.
In order to design concept of the present invention is described, need to analyze neuron action potential forming process further, and when, how neuron at the action potential of diverse location outburst different subtype.Fig. 4 is the structural representation of neuron axon part.The same with Fig. 1, neuron is formed primarily of cell space 1, dendron 2 and aixs cylinder 3.Aixs cylinder 3 can be divided into again initial segment 4 (axon initial segment from cell space to Nerve Terminals In The Human Skin, be called for short AIS), ranvier's constrictions 5 (node of Ranvier), and Nerve Terminals In The Human Skin, initial segment 4 i.e. AIS is divided into again the nearly section 6 of AIS close to cell space, and from the section 7 far away of the AIS away from cell space.Current theory be it is generally acknowledged, although the outburst of neuron action potential can betide any position of neuron in theory, for most neurons, AIS also namely initial segment 4 are trigger regions of action potential.The cell membrane of this position has highdensity valtage-gated Na +ion channel, causes its valve current potential low compared with other positions, so be more easily triggered and snapping motion current potential.From the excitatory signal passed over by cynapse by other neurons of dendron, expand cloth transmission to cell space and aixs cylinder, and integrate at initial segment and AIS, when the current potential of the exciting signal integrated meets or exceeds valve current potential, a large amount of Na +ion channel is opened, and produces action potential outburst.On the one hand to Nerve Terminals In The Human Skin transmission after action potential outburst, act on next neuron, namely form neuronic reflex action; Action potential does back transfer to cell space and dendron on the other hand, acts on the postsynaptic membrane be distributed on cell space and dendron, and therefore produces STDP synaptic plasticity.After action potential outburst, also make membrane depolarization, eliminate residual excitatory potential, (effect of the film discharge circuit namely in neuron analogue means), waits for next time new integration process.
According to current above-mentioned theory, the outburst of neuron action potential, just can back transfer to cell space and dendron, act on the postsynaptic membrane on cell space and dendron, and make these cynapses produce plasticity, comprise cynapse and strengthen LTP and cynapse and suppress LTD.Simply see that this seems rationally meaningful again, according to current " synaptic plasticity theory " about learning and memory, synaptic plasticity is the foundation factors forming memory, and the learning and memory activity of brain relies on synaptic plasticity and realizes.This is also the design theory basis about simulation learning and memory function in the analogue technique of large multi-neuron and neural network at present.
But analyze further, can find that this faces one and makes us the puzzled problem being difficult to again to avoid.According to " the synaptic plasticity theory " of the current memory about study, the acquisition (note) of imformation memory with again obtain (recalling) and employ identical mechanism: the information of input obtains (note) by activating synaptic plasticity and causing the change of synaptic structure to realize to remember, afterwards, again by again activating these synaptic structures to realize the acquisition again (recalling) to imformation memory.So, here Problems existing is: if obtain in the process of (recalling) again at imformation memory, need again to activate these synaptic structures of information of encoding, and cause the process activating again synaptic plasticity, then will inevitably affect the stability of acquired memory, cause the interference to existing memory or destruction.But the fact known to us is really not so: when information input brain, in most of the cases, we just carry out judgement process according to original memory (knowledge, experience, experience, custom etc.) to inputted information, form thinking or action, (even under many circumstances, this think of peace-keeping operations also all unconsciously produce naturally), and not memory process is all carried out to all input information.And if we deliberately do not go change to upgrade original imformation memory, then obtaining again (memory) to imformation memory, can't cause the interference to original memory or destruction.So, also there is theory to think at present, the acquisition (note) of memory and again obtain (recalling) and may have been used different working mechanisms, but employ how different working mechanisms on earth, then not yet have further achievement in research to this.
Recently have research to disclose, action potential has Na +v1.2 and V1.6 two kinds of hypotypes of ion channel, wherein the activation threshold value of V1.6 hypotype action potential is lower, and in AIS section outburst far away, during outburst, action potential is only to Nerve Terminals In The Human Skin unidirectional delivery, carries out signal integration and transmission.The activation threshold value of the action potential of V1.2 hypotype is higher, in cell space outburst, (the V1.2 hypotype action potential that activation threshold value is higher also may break out in the nearly cell space section of AIS, convenient in order to describe, in present specification, for the outburst position of V1.2 hypotype action potential, even if only write cell space, reality also comprises the nearly cell space section of AIS), both to Nerve Terminals In The Human Skin transmission during outburst, also oppositely can change the transmission characteristic of cynapse to cell space and dendron transmission, produce STDP synaptic plasticity phenomenon.But for brain why can with why to form these two kinds different action potential hypotypes, the forming process of the action potential of these two kinds of different subtypes, and corresponding with brain macroscopically physiological function, still there is no result of study.Result of study is had to think due to Na on AIS +the Density Distribution difference of ion channel causes, but the different distributions of this ion channel can be caused why, be that what nervous activity result in this difference, how about the density difference of ion channel causes the formation of two kinds of action potentials to break out, and does not all have explanation at present.
And, according to existing theory, action potential is once in neuronic any position outburst, then will inevitably activate the sodium-ion channel of surrounding, make a large amount of sodium ion inflow, further quickening depolarization, thus make action potential conduct to whole neuron fast along cell membrane, and cause neuron cannot form action potential (refractory period) again due to depolarization.Present Problems existing is: at the film integration process of neuron to input exciting signal, the excitatory potential integrated is once reach the lower activation threshold value of AIS section V1.6 far away hypotype, naturally the first snapping motion current potential here of meeting, and conduct to whole neuron, make whole neuronic membrane depolarization, remove all excitatory potentials, the V1.2 hypotype action potential of high threshold is just had no chance at all and forms outburst.So, why can neuron also at the V1.2 hypotype action potential of cell space or IAS nearly cell space section outburst high threshold? when, neuron just there will be this situation? what physiological phenomenon this macroscopically correspond to again?
Our Water demand brain forms the situation of memory, and namely when clear and definite macroscopically brain just can form memory.According to our real experiences, can find: when information input brain, in most of the cases, brain just carries out reflection process according to original memory to inputted information, forms thinking or action, and not carries out memory process.Brain is general only just forms memory in the following two cases: one, the information of input is enough strong, make we " impression is very dark ", unconscious but remember this information immediately, form memory.This first situation belongs to " passivity memory ", such as, finds a certain fangle, is aware of some unsolved problem answers of a specified duration, or other stimulate very deep information.Two, be that we will go to remember some information in subjective " subconsciousness ", by starting the notice of height, go input (by seeing, read or thinking) and make great efforts to remember information, this needs brain " subconsciousness ", also namely starts intensive concentration and realizes.This second situation belongs to " initiative memory ", the behavior of our daily learning knowledge, and other active memory processes, is all belong to this situation.
So, in conjunction with the forming process of neuron action potential, applicant obtains such analysis result: from the exciting signal of dendron input, be delivered to cell space and IAS section is integrated, in the ordinary course of things, action potential is formed by first triggering in the IAS that activation threshold value is lower section far away, this action potential only does unidirectional delivery to Nerve Terminals In The Human Skin, and carry out exciting signal transmission by presynaptic membrane other neurons forwardly of Nerve Terminals In The Human Skin, this is a kind of signal one-way transmission process, macroscopically show as to input information carry out reactive reflection process.In this case, cell space have high threshold and the lower again V1.2 hypotype sodium-ion channel of channel density, play a kind of inhibition, so that the action potential of IAS section outburst far away can not carry out back transfer to cell space and dendron, so also just do not produce memory.
And neuron wants the action potential that can break out high threshold at cell space or the nearly cell space section of IAS, thus trigger cynapse STDP plasticity and cause memory effect, then there are following two kinds of generation mechanisms: one, because the exciting signal of dendron input is first arrive cell space in transmission, and then arrive aixs cylinder IAS section far away, in this transmittance process, there is certain strength retrogression and time delay in exciting signal, if exciting signal is enough strong, both there is sufficiently high amplitude, (being greater than the activation threshold value of cell space), and there is again enough steep voltage rise slope, (rate of rise k=dv/dt), so that before exciting signal is delivered to aixs cylinder IAS section far away and triggers the action potential forming Low threshold, exciting signal has been integrated in cell space position and has been met or exceeded the threshold potential of cell space, so directly break out the V1.2 hypotype action potential of high threshold on cell space prior to IAS section far away.Because at this moment cell space does not exist inhibition to oneself, so the high threshold action potential that cell space breaks out, just can on the one hand to Nerve Terminals In The Human Skin transmission, produce reflection process, on the other hand also simultaneously to full cell space and dendron back transfer, arrive the postsynaptic membrane on cell space and dendron, enable these cynapses produce synaptic plasticity, thus realize memory effect.It is enough strong that the feature of the first generation mechanism is exactly the exciting signal inputted, and this corresponds to macroscopically our brain and information is produced to the first situation of memory, i.e. " passivity memory ".Two, the second generation mechanism corresponds to macroscopically brain and produces the second situation of memory, i.e. " initiative memory ", is gone the situation carrying out remembering by brain active " subconsciousness ".This be then by neuron between modulation activity realize.When brain carries out " subconsciousness " memory, by the neuron activity of " attention " control path, (" attention " control path of brain is relevant with thalamus and hypothalamus), change the neuronic state of truck, it is made to change to the integration of exciting signal and the activation mechanism of sodium-ion channel, thus at cell space and IAS nearly section snapping motion current potential, cynapse STDP plasticity can be produced.This modulation activity is likely that the neuron of " attention " control path is realized by modularity cynapse.But applicant infers this modulation activity, not follow excitatory signal to be equally realized by chemical synapse, and more may be realized by non-cynapse transmission (contact is transmitted).The brain district of memory effect is easily produced hippocampus etc., (similar also has amygdaloid nucleus, corpus straitum etc.), the arrangement of its cones is extremely tight, rare spongiocyte again therebetween, for electric current provides basis to adjacent cells flowing, when brain starts " subconsciousness " memory, activating appears in some neurocyte of " attention " control path, i.e. snapping motion current potential, makes to occur electric current outside its film.Although this directly can not cause the excited action of the neurocyte of peripheral information path, but these flexible lamp fluid capacitances easily flow to peripheral nerve-cell by space between cells, cause the cell membrane of the cell space of peripheral nerve-cell to occur membrane potential, the activation of sodium-ion channel is made a significant impact.Such as, movable electric current makes the cell membrane of cell space occur membrane potential, because this current potential is depolarization current potential, so directly dragged down the threshold value of the sodium-ion channel of V1.2 hypotype on cell space, or change the activation mechanism of its passage, make its easier snapping motion current potential, thus make neuron directly can break out the action potential of V1.2 hypotype at cell space or the nearly section of IAS, (outburst prior to the V1.6 hypotype action potential of IAS section far away), and back transfer can be carried out to cell space and dendron, the cynapse caused on cell space and dendron produces synaptic plasticity, form memory effect.
To sum up state, neuron is to the integration of dendron input signal and reaction, at aixs cylinder IAS section outburst far away Low threshold action potential in the ordinary course of things, and do unidirectional delivery to Nerve Terminals In The Human Skin, form the reactive reflex activity to input information, this is the process that information obtains (recalling) again, is also the process of human thinking and reflex activity.This process action current potential, not to cell space and dendron back transfer, can not cause the cynapse connected with them and produce synaptic plasticity, so also would not affect the stability of original established memory.And neuron in both cases, by two kinds of different generation mechanisms, (correspond to two kinds of situations that brain forms " passivity memory " and " initiative memory "), the action potential of high threshold just can be broken out at cell space or the nearly section of IAS, both information reflection was formed to Nerve Terminals In The Human Skin transmission, again simultaneously to cell space and dendron back transfer, cause the cynapse connected with them and produce synaptic plasticity, thus form memory.So, the acquisition (note) of imformation memory is with obtaining (recalling) again although be by identical neuronal pathways, but have employed different working mechanisms to realize, the process (being also the process of thinking and other reflex activities) that the information that ensure that obtains again can not be formed original memory to be destroyed and impact, more meets the working condition of brain.
As for why, the V1.6 hypotype action potential of IAS section outburst far away can not carry out back transfer to cell space and dendron, this is likely that inhibitory synapse on cell space is had an effect: when the inhibitory synapse on cell space is movable, cell membrane negative ion (chlorine negative ion) passage on cell space is opened, cause stream in chlorine negative ion, make when the action potential of IAS section outburst far away is close to cell space, the effect that inhibitory synapse produces counteracts the depolarization (in sodium positive ion stream) of action potential, thus hinders action potential to cell space and dendron back transfer.But applicant is more ready to infer, its reason may be that sodium-ion channel exists our unknown space structure and working mechanism.Theory is in the past thought, sodium-ion channel on neuron is independent of one another and presents the same valtage-gated activation mechanism, so once snapping motion current potential, because action potential far exceedes its gate voltage, other sodium-ion channels will inevitably be activated, thus make action potential expand cloth to whole neuron.But applicant thinks, we in the past may be too simple to the understanding of sodium-ion channel: the sodium-ion channel of serotonergic neuron diverse location can have different space structures, cause its activation mechanism significantly different; Or, also may have the same space structure by sodium-ion channel, but show significantly different activation mechanisms because of the modulation of neuromodulator or the impact of bulk potential; Again or, also may be only that different sodium-ion channel has slightly different activation threshold values, but cause the remarkable difference of its activation mechanism.In a word, the different activation mechanisms of sodium-ion channel, cause, at the action potential of aixs cylinder IAS section far away outburst, cannot activating the sodium-ion channel of cell space and the nearly cell space section of IAS, cause action potential cannot to cell space and dendron back transfer.(estimate this be also the exciting signal size of no matter dendron input how, generally all directly cannot trigger at dendron and form action potential, and cell space can only be delivered to or IAS carries out integrating one of reason that could trigger formation action potential).
Applicant, according to above-mentioned research and analysis, for two kinds of these two kinds of action potentials different generation mechanisms and forming process thereof, discloses two kinds of neuron analogue means of the present invention and method thereof.
Fig. 5 is the circuit block diagram of the first neuron analogue means of the present invention.The first neuron analogue means of the present invention, includes dendron input end (dendron for imictron inputs), signal processing module (part integrated signal for imictron and process) and aixs cylinder output terminal (aixs cylinder for imictron exports), as prior art, signal processing module includes film integrating circuit (for simulating the process integrated film input excitatory potential, general normal employing resistance and electric capacity composition RC charging circuit form film integrating circuit), film discharge circuit (after triggering for simulated action current potential film potential depolarization and form the action of refractory period, be also sometimes referred to as depolarization circuit, general normal employing switching elements conductive makes the capacitor discharge of film integrating circuit), (the action of neuron triggering action potentials when the excitement for analogue membrane current potential is integrated and reached activation threshold value of threshold triggers circuit, general normal use voltage comparator circuit.In order to distinguish, first threshold trigger circuit are called in the present invention, its activation threshold value is V1) and action potential pulse-generating circuit (for the output pulse of simulated action current potential, in order to distinguish, the first action potential pulse-generating circuit is called) in the present invention, the general output circuit also including action potential, (there is no substantive analog functuion, just for isolating the action potential pulse signal amplifying action potential pulse-generating circuit and produce); Its concrete annexation is: dendron input end is connected to the input end of film integrating circuit, the output terminal of film integrating circuit connects the input end of first threshold trigger circuit, and the output terminal of first threshold trigger circuit connects the input end of the first action potential pulse-generating circuit; The output terminal of the first action potential pulse-generating circuit is connected to aixs cylinder output terminal by output circuit; The output terminal of the first action potential pulse-generating circuit is connected to film integration discharge circuit simultaneously.
The invention is characterized in: described neuron analogue means also includes Second Threshold trigger circuit, the input end of Second Threshold trigger circuit is connected to the output terminal of film integrating circuit, the output terminal of Second Threshold trigger circuit is connected to the second action potential pulse-generating circuit, the output terminal of the second action potential pulse-generating circuit connects aixs cylinder output terminal, and the output terminal of the second action potential pulse-generating circuit is also connected to dendron input end by a reverse channel transmission simultaneously.The activation threshold value V2 of Second Threshold trigger circuit setting is greater than the activation threshold value V1 of first threshold trigger circuit setting.Its difference is arranged depending on the setting value of V1 and the integration of film integrating circuit, and it is comparatively more suitable than V1 large 10% to 35% to be generally set to.
Above-mentioned neuron analogue means operationally, in usual cases, the signal of dendron input end input exports after the integration integration of film integrating circuit, when output signal voltage slowly rise reach first threshold trigger circuit time, first threshold trigger circuit triggers, make the first action potential pulse-generating circuit output action potential pulse, aixs cylinder output terminal is outputted to by output circuit, realize the integration to input signal and reflection function, and made the capacitor discharge of film integrating circuit by film discharge circuit simultaneously, wait for next time to the integration process of input signal, and drag down integral output voltage, Second Threshold trigger circuit can not be triggered.And when the signal of dendron input end input is very strong, so that the magnitude of voltage of film integrator output signal rises quickly, while first threshold trigger circuit triggers, (or close to simultaneously, first threshold trigger circuit are made also to have little time to make film discharge circuit drag down the output voltage of film integrating circuit), Second Threshold trigger circuit also and then trigger, make the second action potential pulse-generating circuit output action potential pulse, both aixs cylinder output terminal had been outputted to, and simultaneously reverse transfer to dendron input end, form the spike signal required for cynapse STDP plasticity on dendron.
Second Threshold trigger circuit of the present invention are used for the first producing cause and the biological significance thereof that imictron breaks out high activation threshold value action potential, namely cell space or dendron input exciting signal strongly time, the film excitatory potential rate of rise after its integration is very steep, so that directly directly breaks out the action potential of high activation threshold value at cell space or AIS nearly cell space section.Neuron realizes information reflection by the action potential lower at aixs cylinder AIS section far away outburst activation threshold value, by cell space or AIS nearly cell space section outburst activation threshold value higher action potential realize imformation memory, but technology is in the past owing to lacking understanding to the importance that there are these two kinds of different subtype action potentials, particularly neuron is broken out to the essence reason of the action potential of two kinds of different subtypes, and when can break out any hypotype shortage understanding, simulate more accurately so do not do this respect, and in fact how this works in reflection and memory two kinds of different conditions for explaining and simulate same neuron, be even more important.
As a kind of prioritization scheme, between the input end of first threshold trigger circuit and the output terminal of film integrating circuit, be added with a signal delay circuit.Make the output signal of film integrating circuit, be delivered to the time of the input end of first threshold trigger circuit, than the time being delivered to Second Threshold trigger circuit input end, have an of short duration delay.Time delay, value was very little, specifically depending on the time parameter of film integrating circuit, generally, approximately can be taken as integration time constant 5% to 25% between.Signal delay circuit be arranged so that when dendron input end inputs strong signal, (when namely neuron occurs breaking out the first situation of high threshold action potential), Second Threshold trigger circuit can reliably trigger prior to first threshold trigger circuit.When the program is delivered to the AIS of aixs cylinder section far away for simulating from dendron input and the exciting signal integrated, need the time that one of short duration.This of short duration time delay in technology in the past, owing to not understanding its biological meaning, thus not pay attention to by people and simulate.And applicant notices, to the simulation of this time delay, the substantial connection between position and different subtype thereof that the integration transmission of exciting signal of dendron input and action potential break out can be simulated more accurately, more be consistent with the working condition of real cerebral neuron, simulate the operative scenario that cerebral neuron produces memory more accurately.
As another kind optimization, the output terminal of the second action potential pulse-generating circuit is connected to film discharge circuit simultaneously.The work of original film discharge circuit can have been come by the output of the first action potential pulse-generating circuit, but the output terminal of two pulse-generating circuits is controlled film discharge circuit separately, more meet the course of work of neuron two kinds of action potentials, and make to there will not be two kinds of action potential pulses to repeat to produce the situation of superposition.
Further, described neuron analogue means also with one for switching " the attentional manipulation end " of memory/reflective operation state, (be with reflective condition or with remember condition carry out work for directly being controlled by external signal to switch neuron analogue means); Described Second Threshold trigger circuit, with threshold regulating circuit, for turning down the voltage triggered threshold value of Second Threshold trigger circuit, (making the trigger voltage V2 of Second Threshold trigger circuit be less than the trigger voltage V1 of first threshold trigger circuit); The input control end of threshold regulating circuit, is connected to " attentional manipulation end ".During work, " attentional manipulation end " can be connected to other neuronic aixs cylinder output terminals, modulated by other connected neuronic excitation activities, or, directly switch in high level and low level state, the two kinds of duties also namely directly making neuron analogue means process in imformation memory and reflection switch.When " attentional manipulation end " does not have input control signal, neuron is switched in " reflection " duty, carries out work according to working mechanism recited above, and when working as " attentional manipulation end " input effective control signal of appearance, neuron is switched in " memory " duty, the voltage triggered threshold value being turned down Second Threshold trigger circuit by threshold regulating circuit makes it lower than the activation threshold value of first threshold trigger circuit, at this moment when the output signal of film integrating circuit is when constantly integrating rising, Second Threshold trigger circuit will prior to first threshold trigger circuit trigger action, make the second action potential pulse-generating circuit output action potential pulse, both aixs cylinder output terminal had been outputted to, also oppositely output to dendron input end simultaneously, form the spike signal required for cynapse STDP plasticity on dendron.This technical scheme is used for the second producing cause that accurate simulation neuron breaks out the action potential of high activation threshold value, described " attentional manipulation end " is to the adjustment of the setting voltage threshold value of Second Threshold trigger circuit, that simulation brain carries out " subconscious memory " (also i.e. intensive concentration) when working, by the neuronic activation of notice control path, the modulating action that (non-cynapse transmission) is transmitted in contact is created to the cell space of related Neurons or the cell membrane Na ion channel of AIS nearly cell space section, the action potential of high activation threshold value can be broken out at cell space or AIS nearly cell space section, produce action potential reverse transfer phenomenon, thus produce cynapse STDP plasticity, realize the memory function to information.Prior art is owing to lacking theoretic knowledge, so do not have associated analog technology to this Crack cause of outburst high threshold action potential.
As further improvement, described " attentional manipulation end ", also with the delay holding circuit of a duty, keeps certain hour for making duty.This improvement, for simulating the neuron of notice control path, to the modulating action of the neuronic Na ion channel of information processing path, has the maintenance effect of certain hour, more meets the working condition of brain.
Accordingly, the first neuronic analogy method of the present invention, comprising:
(1), integration is carried out to the signal of dendron input end input; (integrating circuit namely corresponding to analogue means carries out integration to the input signal of dendron input end); (2), detect the voltage signal after integration, (being equivalent to employing two threshold triggers circuit to detect the output terminal of integrating circuit); (3) if this voltage signal is less than setting value 1, then neuron does not produce action, (the trigger voltage V1 that the output voltage being equivalent to integrating circuit is less than first threshold trigger circuit also i.e. setting value 1); If this voltage signal is equal to or greater than setting value 1 but its rate of rise is less than setting value 2, (rate of rise k=dv/dt, setting value 2 equals V2/T, wherein V2 is the trigger voltage of Second Threshold trigger circuit, the signal delay time of the signal delay circuit of T added by first threshold trigger circuit input end, general T is 5% to 25% of the integration setting-up time of film integrating circuit.The output voltage that this situation is equivalent to integrating circuit is greater than the trigger voltage V1 of first threshold trigger circuit, but the trigger voltage V2 of Second Threshold trigger circuit cannot be met or exceeded within the signal delay time that setting-up time T is also signal delay circuit), then trigger generation action potential pulse, output to aixs cylinder output terminal, (being equivalent to first threshold trigger circuit triggers generation action potential and Second Threshold trigger circuit do not trigger), the voltage signal values that integration exports is reset simultaneously, if this voltage signal is equal to or greater than setting value 1 and its rate of rise is equal to or greater than setting value 2, (output voltage being equivalent to integrating circuit is greater than the trigger voltage V1 of first threshold trigger circuit, and within setting-up time T, meet or exceed the trigger voltage V2 of Second Threshold trigger circuit), then trigger generation action potential pulse, not only output to aixs cylinder output terminal but also oppositely output to dendron input end simultaneously, (be equivalent to Second Threshold trigger circuit triggers and produce action potential, not only be transported to aixs cylinder output terminal but also be transported to aixs cylinder input end by reverse channel transmission) simultaneously, the voltage signal values that integration exports is reset simultaneously.
Further, neuron analogy method of the present invention also switches memory and reflection two kinds of different operating states by the signal condition of " attentional manipulation end ", (be with reflective condition or with remember condition carry out work for being controlled by external control signal to switch neuron analogue means), when " attentional manipulation end " signal condition is " reflection " state, neuron carries out work by aforementioned analog method, when " attentional manipulation end " signal condition is " memory " state, then by above-mentioned the (3) step change into: if this voltage signal is less than setting value 1, then neuron does not produce action, if this voltage signal is equal to or greater than setting value 1, then trigger and produce action potential pulse, not only output to aixs cylinder output terminal but also oppositely output to dendron input end simultaneously, (be equivalent to " attentional manipulation end " turns down Second Threshold trigger circuit trigger voltage V2 by threshold regulating circuit, V2 is made to be equal to or slightly less than V1, at this moment when integral voltage reaches V1, Second Threshold trigger circuit have just triggered generation action potential, not only be transported to aixs cylinder output terminal but also be transported to aixs cylinder input end by reverse channel transmission) simultaneously, the voltage signal values that integration exports is reset simultaneously.
Fig. 6 is that a kind of single-chip microcomputer that adopts is to realize the physical circuit of the neuron analogue means shown in Fig. 5.Wherein R601, C601 form the film integrating circuit to dendron input end input signal, and C601 is film integrating capacitor; MCU601 is single-chip microcomputer, and wherein I/O1 is the input end of first threshold trigger circuit, and R602, C602 form the delay circuit of its input, and I/O3 is the output terminal of the first action potential pulse-generating circuit; I/O2 is the input end of Second Threshold trigger circuit, and I/O4 is the output terminal of the second action potential pulse-generating circuit; Triode T601 forms film discharge circuit, discharges during conducting to C601; T602, T603 form the output circuit of action potential; T604, T605 accept the output of I/O4, oppositely output to dendron input end after amplification; C603, R603 form the delay holding circuit of " attentional manipulation end " input signal, and be input to I/O5, when input reaches certain level magnitudes (effectively), adjusted by the activation threshold value of threshold regulating circuit to Second Threshold trigger circuit of the inside.MCU601 is by its program, according to principle of work and the neuron analogy method of the present invention of the neuron analogue means of Fig. 5, realize generation and the logic function of its threshold triggers and action potential, this belongs to these those skilled in the art without the need to creativeness just attainable technology.In order to accompanying drawing is succinct, that Fig. 6 does not mark other necessity of single-chip microcomputer yet but belong to the peripheral circuit of routine techniques.
Fig. 7 is the another kind of circuit theory diagrams of the neuron analogue means realizing Fig. 5, and this circuit is without the need to adopting single-chip microcomputer.Wherein R701, C701 form the film integrating circuit to dendron input end input signal, and C701 is film integrating capacitor; R702, C702 form the delay circuit of input; IC701, IC702 (voltage comparator) form the first and second threshold triggers circuit; T703, IC703 (555 time base circuit) and T704, IC704 form the first and second action potential pulse-generating circuits; T705, T706 form the output circuit of action potential pulse; T701 forms film discharge circuit; C703, R703 form the delay holding circuit of " attentional manipulation end " input signal, dragged down the magnitude of voltage of the inverting input reference voltage of Second Threshold trigger circuit and IC702, also namely drag down the activation threshold value of IC2 when having an input signal by T702 conducting.
Fig. 8 is the circuit block diagram of the second neuron analogue means of the present invention.This neuron analogue means includes dendron input end (for simulating dendron), signal processing module (part integrated signal for imictron and process) and aixs cylinder output terminal (for simulating aixs cylinder) equally, as prior art, signal processing module includes film integrating circuit (for simulating the process integrated film input excitatory potential), film discharge circuit (after triggering for simulated action current potential film potential depolarization and form the action of refractory period), (the action of neuron triggering action potentials when the excitement for analogue membrane current potential is integrated and reached threshold value of threshold triggers circuit, in order to distinguish, be called first threshold trigger circuit in the present invention, its activation threshold value is V1), with the first action potential pulse-generating circuit (the output pulse for simulated action current potential), the general output circuit also including action potential, for isolating the action potential pulse that amp pulse generation circuit produces, its concrete annexation is: dendron input end is connected to the input end of film integrating circuit, the output terminal of film integrating circuit is connected to the input end of first threshold trigger circuit, and the output terminal of first threshold trigger circuit is connected to the input end of the first action potential pulse-generating circuit, the output terminal of the first action potential pulse-generating circuit connects aixs cylinder output terminal, the output terminal of the first action potential pulse-generating circuit is connected to film discharge circuit simultaneously, technical characteristic of the present invention is: described neuron analogue means is also with action potential reverse channel transmission, the output terminal of the first action potential pulse-generating circuit, is connected to dendron input end by this reverse channel transmission, the break-make of reverse channel transmission is subject to the control of a reverse transfer control circuit, the control input end of reverse transfer control circuit, is connected to one for switching " the attentional manipulation end " of memory/reflective operation state.
The work of this neuron analogue means, adopt " attentional manipulation end " directly between imformation memory and reflection process two kinds of duties, to carry out switching to neuron analogue means to control, that is to say when " attentional manipulation end " input is effective, when being namely in " memory " state, by directly opening reverse transfer control circuit, make when neuron snapping motion current potential, action potential can carry out reverse transfer by reverse channel transmission to dendron end simultaneously, thus enable the synapse device be connected on dendron produce cynapse STDP plasticity, to realize the memory function to information.And work as " attentional manipulation end " when being in " reflection " state, close reverse transfer control circuit, action potential only exports to aixs cylinder and does not have reverse transfer, thus only produces the integration of input signal and reflection function and do not have memory function.
As optimization, between " attentional manipulation end " and the control input end of reverse transfer control circuit, being also provided with the delay holding circuit of a duty, keeping certain hour for making duty.The program, for simulating the neuron of notice control path, to the modulating action of the neuronic Na ion channel of information processing path, has the maintenance effect of certain hour.
Above-mentioned the second neuron analogue means, directly control neuronic action potential reverse channel transmission by attentional manipulation end, be the situation that imictron the second produces action potential reverse transfer, namely brain realizes memory action by " subconscious memory " (intensive concentration).Certainly on the basis of this, also can more simultaneously imictron another kind produce the situation of action potential reverse transfer, namely the exciting signal of dendron input strongly time, directly cause neuron to break out the action potential of high activation threshold value, and to cell space and dendron reverse transfer.Simulate this working condition, neuron analogue means can be made to include Second Threshold trigger circuit again, the input end of Second Threshold trigger circuit is connected to the output terminal of film integrating circuit, and the output terminal of Second Threshold trigger circuit is connected to the control input end of reverse transfer control circuit.The high threshold action potential that Second Threshold trigger circuit break out for simulating cell space or the nearly cell space section of ISA, so the activation threshold value V2 of Second Threshold trigger circuit setting is greater than the activation threshold value V1 of first threshold trigger circuit setting.General V2 is comparatively more suitable than V1 large 10% to 35%.
Same, as the one optimization of such scheme, in this case, preferably between the input end and the output terminal of film integrating circuit of first threshold trigger circuit, be provided with a signal delay circuit.The program is delivered to the of short duration time delay of a section of the AIS section far away of aixs cylinder for simulating from the exciting signal of dendron input, existingly above its technique effect and meaning describes.
As another kind optimization, between the output terminal and the control input end of reverse transfer control circuit of Second Threshold trigger circuit, be also provided with a Time delay holding circuit.Its work is: when the output terminal of Second Threshold trigger circuit produces output signal, Time delay holding circuit is triggered and exports control signal and make output signal keep the certain time-delay time, thus makes reverse transfer control circuit open conducting and keep conducting state within a certain period of time by output signal.The pulse width of the action potential that Time delay time of Time delay holding circuit should produce slightly larger than action potential pulsing circuit is to make the reverse transfer process of action potential pulse, more reliably complete.
Correspondingly, the second neuron analogy method of the present invention, comprising:
(1), integration is carried out to the signal of dendron input end input; (integrating circuit namely corresponding to analogue means carries out integration to the input signal of dendron input end); (2), detect the output voltage signal after integration, (being equivalent to the output end signal that first threshold trigger circuit detect integrating circuit), if this voltage signal is less than setting value 1, then do not produce action, (output voltage being equivalent to integrating circuit is less than the trigger voltage V1 of first threshold trigger circuit, and at this moment first threshold trigger circuit do not have trigger action); If the voltage signal (3) after integration is greater than setting value 1, (output voltage being equivalent to integrating circuit is greater than the trigger voltage V1 of first threshold trigger circuit), then detect the signal condition of " attentional manipulation end ";
If " attentional manipulation end " signal condition is " reflection " state (being such as low level 0), then neuron carries out work in " reflection " mode, neuron triggers generation action potential pulse, output to aixs cylinder output terminal, (be equivalent to first threshold trigger circuit triggers, first action potential pulsing circuit produces action potential, and is sent to aixs cylinder output terminal), the voltage signal values after integration is reset simultaneously; (being equivalent to the action of film discharge circuit, the capacitor discharge by film integrating circuit);
If " attentional manipulation end " signal condition is " memory " state (such as high level 1), then neuron works in " memory " mode, (be equivalent to reverse transfer control circuit and open conducting), neuron triggers generation action potential pulse, both output to aixs cylinder output terminal, oppositely output to dendron input end again; (be equivalent to first threshold trigger circuit triggers, first action potential pulse-generating circuit produces action potential, both be sent to aixs cylinder output terminal, be sent to dendron input end by reverse transfer circuit again simultaneously), the voltage signal values after integration is reset simultaneously; (being equivalent to the action of film discharge circuit, by film integrating circuit capacitor discharge).
Same conduct one is optimized, described " attentional manipulation end ", also with the delay holding circuit of a duty, for making duty keep certain hour, at this moment the signal condition detecting " attentional manipulation end " is changed into the state of the output terminal of testing lag holding circuit.This improvement, for simulating the neuron of notice control path, to the modulating action of the neuronic Na ion channel of information processing path, has the maintenance effect of certain hour.
Above-mentioned the second neuron analogy method, directly control to switch neuronic working method by " attentional manipulation end ", be the situation of the second generation action potential reverse transfer of imictron, namely brain realizes memory action by " subconscious memory " (being also intensive concentration).On the basis of this, also can more simultaneously imictron another kind produce the situation of action potential reverse transfer, namely the exciting signal of dendron input strongly time, directly cause neuron to break out the action potential of high activation threshold value, and to cell space and dendron reverse transfer.At this moment its analogy method also comprises further:
(4), detect the voltage signal after integration, if this voltage signal its voltage rise slope while being greater than setting value 1 is also greater than setting value 2, (voltage rise slope k=dv/dt, setting value 2 equals V2/T, wherein V2 is the activation threshold value of Second Threshold trigger circuit, the signal delay time of the signal delay circuit of T added by first threshold trigger circuit input end.The output voltage that this situation is equivalent to integrating circuit is greater than the trigger voltage V1 of first threshold trigger circuit, and within setting-up time T, meet or exceed the trigger voltage V2 of Second Threshold trigger circuit), neuron is then made to carry out work in " memory " mode, (be equivalent to Second Threshold trigger circuit and export trigger pip, reverse transfer control circuit is made to open conducting), neuron triggers generation action potential pulse, both output to aixs cylinder output terminal, oppositely output to dendron input end again; (be equivalent to the first action potential pulsing circuit and produce action potential, be both sent to aixs cylinder output terminal, and be sent to dendron input end by reverse transfer circuit again simultaneously), the voltage signal values after integration is reset simultaneously; (being equivalent to the action of film discharge circuit, the capacitor discharge by film integrating circuit).
Fig. 9 is that a kind of single-chip microcomputer that adopts is to realize the physical circuit of the neuron analogue means shown in Fig. 8.Wherein R901, C901 form the film integrating circuit to dendron input end input signal, and C901 is film integrating capacitor; MCU901 is single-chip microcomputer, and wherein I/O1 is the input end of first threshold trigger circuit, and R902, C902 form the delay circuit of its input, and I/O3 is the output terminal of the first action potential pulse-generating circuit; I/O2 is the input end of Second Threshold trigger circuit, and I/O4 is the output terminal of Time delay holding circuit; T901 forms film discharge circuit; T905, T906 form the output circuit of action potential pulse; C903, R903 form the delay holding circuit of " attentional manipulation end " input signal; T904 forms reverse transfer control circuit, for carrying out switch control rule to the action potential signal of the reverse channel transmission from output circuit; The output that T902 accepts the control of " attentional manipulation end ", T903 accepts I/O4 controls, and the break-make of control T904 simultaneously.
Figure 10 is the another kind of circuit theory diagrams realizing Fig. 8 neuron analogue means, and this circuit is without the need to adopting single-chip microcomputer.Wherein R101, C101 form film integrating circuit; R102, C102 form the delay circuit of input; IC101, IC102 (voltage comparator) form the first and second threshold triggers circuit; T103, IC103 (555 time base circuit) form the first action potential pulse-generating circuit; T104, IC104 form Time delay holding circuit; T105, T106 form the output circuit of action potential pulse; T101 forms film discharge circuit; T107 forms reverse transfer control circuit; C103, R103 form the delay holding circuit of " attentional manipulation end " input signal, make T107 conducting when having signal to input by T102.T102 also by the control of IC104, when it has output, makes T107 conducting by T102 simultaneously equally, is also the reverse channel transmission of opening action current potential.
And as the further improvement to two kinds of above-mentioned neuronic analogue means of the present invention or analogy method, described neuron analogue means is all also with modularity synaptic input circuit and modularity synaptic input end; Modularity synaptic input end is connected to the input end of modularity synaptic input circuit, and the output terminal of modularity synaptic input circuit is connected to the output terminal of film integrating circuit.Described modularity synaptic input circuit comprises two aspects: enhancement synaptic input circuit and enhancement synaptic input end, inhibitory synapse input circuit and inhibitory synapse input end thereof.These modularity input circuits make neuron analogue means except having the dendron input end for excitatory signal input, also have " aixs cylinder-aixs cylinder " the type synaptic input function for carrying out modulating action to information processing.Notice that modularity input signal of the present invention is not connected on the input end of film integrating circuit, so it participate in directly unlike the excitatory signal of dendron input film potential integration and can direct triggering action potentials, but be connected on the output terminal of film integrating circuit, form a modulation voltage and enhancer or inhibitor is carried out to film integral output voltage, by the impact on film integral output voltage, can the exciting signal from dendron input and integration be modulated by triggering action potentials.This function is conducive to building the Complex Neural Network that hyperchannel multi information is intermodulated, to simulate experience or mood (happy, frightened, reward with bounties, the habituation etc.) impact on neuronal messages process.
Figure 11 is the circuit theory diagrams of modularity input circuit.Circuit only draws the modulation circuit part of increase, namely this part be the neuron analogue means rising to Fig. 6, Fig. 7, Fig. 9 or Figure 10 circuit on, wherein film integrating capacitor CX is also the integrating capacitor in the film integrating circuit of each circuit, as the C601 of Fig. 6.The signal of enhancement modulation input end is input to the input end of reverser F111 (digital gate circuit) after electric resistance partial pressure, when input signal is effective, reverser F112 positive output, after resistance R111 and R112 dividing potential drop, charged to electric capacity C111 by D111, electric capacity two ends are made to reach certain voltage (generally can be 1/to two/3rd of first threshold trigger circuit triggers voltage), this voltage releases (namely the time of releasing also approximate greatly the effective time of modulation) by R113 on the one hand, be added on film integrating capacitor CX by R114 and D112 on the one hand, film integrating capacitor is made to have certain initial voltage, and when the voltage of film integrating capacitor exceedes initial voltage, because the single effect of diode D112 can not have an impact to the voltage of film integrating capacitor again.Like this, when its dendron input end input pulse signal, the voltage in film integrating capacitor more easily reaches the trigger voltage of threshold triggers circuit, produces action potential, also namely achieve the object of enhancement modulation so more easily trigger.The operation class of inhibition modulation seemingly, the lower voltage just electric capacity C112 produced, acting in opposition due to D114 produces initial voltage can't to film integrating capacitor, and when the voltage of film integrating capacitor is higher, shunting is formed on the contrary by D114 and R118, make the voltage of film integrating capacitor be not easy to reach trigger voltage, threshold triggers circuit is not easy to trigger, and reaches the object of inhibition modulation.
Coordinate the cynapse analogue means of work of the present invention, the various analogue means fully can simulating the operating characteristic of chemical synapse in prior art can be adopted, also can adopt the cynapse analogue technique that below the present invention is disclosed.
Applicant researchs and analyses, in cerebral nerve network, the nerve synapse that at least existence three kinds is dissimilar: the first belongs to more stable cynapse, it transmits usefulness and only has journey synaptic plasticity in short-term, comprise synaptic facilitation (synaptic facilitation), cynapse suppresses (synaptic depression), tetanic rear enhancing (posttetanic potentiation, i.e. PTP) phenomenon, but do not embody long term synaptic plasticity, (it also can realize synaptic plasticity in other words, but under current duty, do not embody synaptic plasticity), they are present in importing into of various signal, export and (such as in unconditioned reflex loop) in other direct and stable signal transmission paths, transmit usefulness comparatively strong and stable, when presynaptic neuron activates, often postsynaptic neuron can be made to trigger generation action potential by strong effective usefulness of transmitting, (some can also be monosynaptic transmission, namely the transmission of signal of single cynapse is just enough to activate postsynaptic neuron), the present invention is referred to as here " stablizing cynapse ".The second has obvious long term synaptic plasticity, comprise long term potentia ̄tion (long-term potentiation, i.e. LTP) phenomenon and long-term depression (long-term depression, i.e. LTD) phenomenon, according to current research, cynapse Long-term plasticity has the Temporal dependency characteristic of presynaptic membrane and cynapse spike below, (Spike Timing-Dependent Plasticity, i.e. STDP).This kind of cynapse is extensively present in hippocampus, corpus straitum, amygdaloid nucleus and Brain cortex, significant for the study and long-term memory that realize information, maximum for the analogue technique disclosed in this class cynapse at present, the present invention is herein referred to as " modifiable synapse ".The third cynapse is never familiar with and pay abundant attention, so also not relevant analogue technique.Although this kind of cynapse has generated the physiology and appearance possessing cynapse, but temporarily also do not possess normal cynapse and transmit usefulness, need neuronic repeatedly association after the presynaptic after stimulating, could start and be formed effective cynapse and transmit usefulness, (this wherein should also need promotor gene to transcribe and protein synthesis), they are at least be present in cortex to be especially present on the neocortex of process high-level information, significant for the long-term memory realizing declarative information.And, applicant infers from the macroscopic token of imformation memory further, even started and formed effective cynapse to transmit usefulness, if do not stimulated for a long time, (do not use in thinking or again do not remember), these cynapses still may lose some cynapse material (such as protein) due to the competition of other cynapses, cause cynapse again to be lost efficacy.This third cynapse is called " cynapse of prefabricated property " by the present invention.About these three kinds of cynapses, especially wherein " modifiable synapse " and " cynapse of prefabricated property " not same-action in memory, long-term memory and permanent memory and transforming mutually, can with reference to below " three, working memory, long-term memory (hippocampus memory) and (forever) difference of remembering and conversion for a long time " content partly.
For above-mentioned three kinds of cynapses, the present invention adopts three kinds of technology to realize the simulation of these cynapses.
The circuit of " solidification cynapse " of the present invention analogue means is as Figure 12." solidification cynapse " analogue means comprises cephacoria input end, caudacoria output terminal, it is characterized in that; Between cephacoria input end and caudacoria output terminal, be serially connected with a capacitor discharging circuit, capacitor discharging circuit simultaneously and be connected to an electric capacity leadage circuit.Formed with this and the most simply be there is the cynapse transmission channels that usefulness is transmitted in fixing cynapse, also namely do not possessed the cynapse transmission channels of synaptic plasticity.As the simplest scheme of one, capacitor discharging circuit is made up of a discharge capacity C801, and the front end of discharge capacity connects cephacoria input end, and the rear end of discharge capacity is transmitted resistance R801 by one and connected caudacoria output terminal; Electric capacity leadage circuit is used for the electric charge of discharge capacity slowly to release, and described electric capacity leadage circuit comprises bleeder resistance R802, a R803 and diode D801; Bleeder resistance R802 is connected between the front end of discharge capacity C801 and ground (i.e. the common ground end of circuit); Diode D801 is oppositely connected between the rear end of discharge capacity and ground after being connected in series another resistance R803.Diode D802 is used for carrying out unidirectional isolation to input, and synaptogenic signal intensity can not be impacted other circuit before cephacoria input end.Transmit resistance R801 for setting the size (weight) of the transmission usefulness of this cynapse, it is different for transmitting usefulness in its cynapse of information processing passage that brain is different, such as larger in the transmission usefulness of its cynapse of input and output loop of information, even single cynapse simple venation flushing is enough to next neuron fires is broken out, and it is less to transmit usefulness in the cynapse of intrerneuron, general need multiple cynapse to carry out spatial integration or single cynapse time of carrying out to integrate and could cause next neuron snapping motion current potential.On resistance 801, reverse parallel connection diode D803, D803 are used for providing one oppositely to the particular path of electric capacity C801 reverse charging to the action potential pulse from postsynaptic membrane output terminal.Change resistance R802 and R803 and can change capacitor discharge speed, also namely change cynapse suppresses phenomenon suppression degree to low frequency action potential.During work, if the input of cephacoria input end is the pulse of low frequency action potential, due to the effect of electric capacity C801, after each action potential pulse is passed through, need the leadage circuit that a period of time allows electric capacity consist of diode and resistance and discharge, could recover.If continuous two or several low-frequency pulse, because electric capacity is had little time electric discharge by charging, larger impedance can be presented to the pulse of second immediately, even if also cynapse transmission produces the phenomenon of cynapse suppression to low frequency action potential.If the input of cephacoria input end is high frequency spike pulse, because the impedance of electric capacity to high-frequency impulse is less, so impact is less.If particularly high frequency mo current potential can cause the neuronal integration of postsynaptic membrane to break out the V1.2 hypotype action potential with high threshold, then because action potential meeting back transfer is to postsynaptic membrane, electric capacity C801 reverse charging is made by diode D803 and R802, thus make C801 present the transmission usefulness obviously strengthened to next action potential pulse, that is to say that making cynapse transmission produce cynapse to high frequency mo current potential strengthens phenomenon.
" solidification cynapse " of the present invention analogue means adopts extremely simple circuit to realize the cynapse propagation function of common cynapse, although strengthen just a kind of simply rough simulation to cynapse suppression and cynapse, but it is extremely simple to have circuit, thus the feature that cost and volume are extremely low is also had, can be applicable to the signal input and output path in simulative neural network, on the direct conduction paths of signal such as signal feedback path, obviously can save cost when simulative neural network larger.
For cynapse STDP plasticity, current theory is thought: the STDP characteristic of synaptic plasticity, also the transmission usefulness of i.e. cynapse, sequential with action potential spike neuronic after the presynaptic is correlated with, as Figure 13, if the time tpost that the time tpre that the spike Vpre of presynaptic neuron produces produces early than the spike Vpost of postsynaptic neuron, also be △ t=(tpost-tpre) > 0, then produce synaptic trasmission (long-term potentiation, i.e. LTP) phenomenon, and the delay of spike and the less then cynapse of △ t are transmitted, and to strengthen LTP effect larger, (for the ease of describing, this situation, we can call that its spike signal has LTP feature), if instead the spike of presynaptic neuron is slower than the spike of postsynaptic neuron, then produce cynapse and transmit long-term depression (long-term depression, i.e. LTD) phenomenon, and the delay of spike is less, it is larger that cynapse transmission weakens LTD effect, (for the ease of describing, this situation, we can call that its spike signal has LTD feature).The LTP characteristic sum LTD feature of spike signal, all belongs to the STDP plasticity that cynapse is transmitted.
Have plastic cynapse analogue means in prior art and disclose multiple technologies, comprise up-to-date employing memory varistor material and directly manufacture modifiable synapse, this is obviously more convenient.And adopt electronic circuit to simulate the basic circuit of modifiable synapse as Figure 14, generally comprise cephacoria input end, caudacoria output terminal, STDP chronicle recognition circuit, usefulness Circuit tuning is transmitted in STDP computing circuit and cynapse, the spike signal that STDP chronicle recognition circuit can occur according to cephacoria input end and caudacoria output terminal, identify and judge whether its spike signal has STDP feature, and the LTP feature belonged to wherein or belong to LTD feature, the signal of these two kinds of features, be output to a STDP computing circuit and carry out plus-minus accumulation, output numerical value after computing transmits usefulness Circuit tuning for controlling cynapse, by transmitting usefulness Circuit tuning, the numerical value (that is to say the power that cynapse outputs signal) that usefulness is transmitted in cynapse is adjusted, change with the plasticity that this realizes transmitting cynapse output signal, namely synaptic plasticity is realized.Prior art adopts microprocessor to realize the above-mentioned function comprising STDP chronicle recognition circuit, STDP computing circuit and cynapse transmission usefulness Circuit tuning by program mostly.And the present invention discloses another kind of analogy method and device.
With reference to Figure 15.The analogy method that the present invention has " modifiable synapse " of cynapse transmission STDP plasticity function comprises:
(1), controlled the output signal of caudacoria output terminal by the input signal of presynaptic membrane input end, make output signal synchronous with input signal; (2), detect the spike signal of presynaptic membrane input end and caudacoria output terminal; (3), when recognizing the spike signal occurring meeting STDP characteristic between cephacoria input end and caudacoria output terminal, then by a charge circuit and a discharge loop, discharge and recharge is carried out to an accumulate device according to the feature of signal: if signal is LTP feature, then charge, if signal is LDP feature, then discharge; The time width of charge or discharge is determined by STDP characteristic; (4), with the magnitude of voltage of accumulate device, adjust the size of voltage (or electric current) amplitude of cynapse output signal, and output to caudacoria output terminal; (5), accumulate device is set up the initialization voltage with certain numerical value when starting working.
As a kind of prioritization scheme, when the voltage on accumulate device is greater than or less than initialization voltage, slowly eliminate this voltage difference by a two-way leadage circuit, until the voltage on accumulate device equals initialization voltage.(program simulation to the slow elimination of cynapse LTP and LTD effect, make cynapse analogy method of the present invention more perfect).
As Figure 15.The analogue means that the present invention has " modifiable synapse " of cynapse transmission STDP plasticity function comprises cephacoria input end, caudacoria output terminal, STDP chronicle recognition circuit and transmits usefulness Circuit tuning.STDP chronicle recognition circuit belongs to prior art, the spike signal that can occur according to cephacoria input end and caudacoria output terminal, identifies and judges whether its spike signal has STDP feature, and identification is judged belong to LTP feature or belong to LTD feature.In addition, STDP chronicle recognition circuit all has magnitude of voltage to the signal that cephacoria input end and caudacoria output terminal input and detects, magnitude of voltage due to the spike of action potential is obviously greater than the signal voltage value after cynapse is transmitted, so a STDP chronicle recognition circuit action potential spike higher to voltage amplitude identifies, and the cynapse output signal of other synapse device cynapse that cynapse self produces being outputed signal and passes over from caudacoria output terminal, can not be identified because signal voltage amplitude is lower.STDP chronicle recognition circuit can adopt the microprocessors such as CPU or MCU, carries out analog to digital conversion to the spike signal of cephacoria input end and caudacoria output terminal, then according to its sequential priority and mistiming, identifies and export LTP characteristic signal and LTD characteristic signal.Also voltage comparator and time base circuit be can adopt, according to the sequential of spike signal successively and the mistiming, LTP characteristic signal and LTD characteristic signal exported by logic identification.This belongs to prior art, has cynapse and transmits in the cynapse analogue technique that STDP plastic employing electronic circuit realizes disclosed and have correlation technique.And, the work characteristics about STDP chronicle recognition circuit described here and actualizing technology, all equally applicable in other parts of the present invention.
It is characterized in that: STDP chronicle recognition circuit is with LTP output terminal and LTD output terminal; (there is the priority sequential control LTP output terminal of spike and the output signal of LTD output terminal according to cephacoria input end and caudacoria output terminal in described STDP chronicle recognition circuit); LTP output terminal and LTD output terminal are connected to an accumulate device for store electrical energy respectively by a charge circuit and a discharge loop.STDP chronicle recognition circuit LTP output terminal when identifying LTP feature occurs that LTP pulse exports, and the time interval of the pulse width of LTP pulse and former and later two spike signals presents negative correlation, also STDP characteristic is namely met, LTP pulse is charged to accumulate device by charge circuit, make the voltage rise at accumulate device two ends, (representing the lifting that usefulness is transmitted in cynapse); STDP chronicle recognition circuit LTD output terminal when identifying LTD feature occurs that LTD pulse exports, and the time interval of the pulse width of LTD pulse and former and later two spike signals presents negative correlation, LTD pulse makes accumulate device discharge by discharge loop, accumulate device both end voltage is declined, (representing the decline that usefulness is transmitted in cynapse).Accumulate device can adopt the capacitor that a leakage current is very little, or the electricity memory device of other low capacities.
Part is connected with an initialization charging circuit; (initialization charging circuit is used for making accumulate device be charged to initialization voltage rapidly when circuit is started working, possesses initial quantity of electricity, cynapse is made just to possess the transmission usefulness of " effectively " in original state, its initialization voltage also i.e. reference voltage of accumulate device when not there is synaptic plasticity, usefulness is transmitted in the cynapse that at this moment cynapse exports as standard).As a kind of concrete scheme, initialization charging circuit comprises a mu balanced circuit, the charging output voltage of mu balanced circuit equals the initialization voltage of accumulate device, mu balanced circuit is connected to accumulate device by a unidirectional rapid charge loop, when circuit is started working, mu balanced circuit to the rapid charge of accumulate device, makes the voltage of accumulate device reach initialization voltage fast by rapid charge loop, i.e. reference voltage, rapid charge loop is closed afterwards.
The output of accumulate device is connected to cynapse and transmits usefulness Circuit tuning, and as the reference voltage that it exports, the synchro control that usefulness Circuit tuning accepts the input signal of cephacoria input end is simultaneously transmitted in cynapse.Like this, the output signal transmitting usefulness Circuit tuning is subject to the modulation of accumulate device output voltage in amplitude, the size of accumulate device output voltage determines the size of output signal voltage amplitude, also namely determine that cynapse exports the size of usefulness, and in time by the synchro control of the input signal of cephacoria input end; The output signal transmitting usefulness Circuit tuning exports as cynapse signal and is connected to caudacoria output terminal.Its output terminal can also include the circuit of above-mentioned " solidification cynapse " analogue means, makes it the transmission characteristic with cynapse enhancing and cynapse suppression.
The analogy method of " modifiable synapse " of the present invention and device, simulate the transmission characteristic of the chemical synapse with synaptic plasticity, but it adopts and changes the method that circuit transfer impedance (variable resistance circuit) realizes transmitting usefulness unlike prior art, but identify LTP and LTD feature by STDP chronicle recognition circuit, and by charging circuit and discharge circuit, discharge and recharge is carried out to an accumulate device, the voltage on accumulate device is made to produce height change along with charging (during LTP) and electric discharge (during LTD), and for modulating the output of synaptic efficacy Circuit tuning, thus the magnitude of voltage making cynapse export pulse produces the corresponding adjustment of height, the simulation to STDP synaptic plasticity is realized with this.
As one optimization, accumulate device is also connected with a two-way leadage circuit; Two-way leadage circuit is used for slowly eliminating accumulate device causes voltage rising or reduction change due to discharge and recharge within a certain period of time, until the voltage on accumulate device equals initialization voltage.Thus the slow elimination of simulating cynapse LTP and LTD effect, make cynapse analogue means of the present invention more perfect, the analog operation of continuous multi-process can be adapted to.As the simple scheme cleverly of one, two-way leadage circuit comprises the mu balanced circuit that has setting voltage, (setting voltage of mu balanced circuit equals the initialization voltage of accumulate device, also be the reference voltage of accumulate device when not there is plasticity, at this moment the transmission usefulness for standard is transmitted in cynapse), mu balanced circuit is connected to accumulate device by a resistance, when the voltage of accumulate device is greater than the setting voltage of mu balanced circuit, when also namely there is cynapse LTP effect, accumulate device is slowly discharged to mu balanced circuit by resistance, also namely slowly eliminates LTP effect; When the voltage of accumulate device is less than voltage of voltage regulation setting voltage, when also namely there is cynapse LTD effect, mu balanced circuit is slowly charged to accumulate device by resistance, also namely slowly eliminates LTD effect.After a certain time, after the term of validity of also i.e. synaptic plasticity, make the voltage on accumulate device equal the setting voltage of mu balanced circuit, that is to say normal output pulse amplitude when plasticity does not appear in cynapse.
Figure 16 is a kind of circuit realizing Figure 15 " modifiable synapse " analogue means.Wherein MCU1601 is for identifying STDP spike and exporting LTP and LTD signal.Signal from cephacoria input end and caudacoria output terminal is input to I/O1 and I/O2, MCU1601 is according to the sequential relationship of cephacoria input end and caudacoria output terminal two signals, identify and whether belong to LTP effect or STD effect, and the intensity of effect, (according to Figure 13), and exported by I/O3 and I/O4.When there is LTP effect, I/O3 exports positive level, the intensity positive correlation of its pulse width exported and LTP effect, and is charged to accumulate device and electric capacity C1601 by D1602 and R1602, makes its voltage rise; When there is LTD effect, I/O4 exports negative level, the intensity positive correlation of its pulse width exported and LTD effect, and is discharged to accumulate device and electric capacity C1601 by D1601 and R1601, makes its voltage drop; And when there is not LTP effect or STD effect, then I/O3 and I/O4 does not effectively export, (unsettled or output opposite levels).Like this, the voltage on electric capacity C1601 along with cephacoria input end and caudacoria output end signal STDP spike LTP and LTD effect and there is respective change.
T1601, D1603, C1602 etc. form initialization charging circuit.Wherein D1603 forms mu balanced circuit, and voltage stabilizing output voltage equals the initialization voltage of accumulate device C1601, is also reference voltage; T1601, C1602 etc. form unidirectional rapid charge loop and are connected to C1601, and when circuit starts to be energized work to C1601 rapid charge, make its voltage reach reference voltage fast, rapid charge loop is closed afterwards, until circuit restarts next time.R1604, D1604, R1605 etc. form two-way leadage circuit; Wherein R1604, D1604 form mu balanced circuit, and voltage of voltage regulation value equals reference voltage, and is connected to accumulate device C1601 by R1605; When the voltage of C1601 is greater than reference voltage, then C1601 is slowly discharged to mu balanced circuit by R1605, slowly eliminates LTP effect; When the voltage of C1601 is less than reference voltage, mu balanced circuit is slowly charged to C1601 by R1605, also namely slowly eliminates LTD effect.
The voltage of accumulate device C1601, isolated by the amplifier IC1601 forming follower circuit, be connected on analog switch K1601, the switch of K1601 controls by the signal of cephacoria input end, when cephacoria input end occurs that pulse signal is also action potential signal, K1601 synchronously opens, output sync pulse signal, the voltage amplitude of its pulse signal equals the voltage amplitude of C1601, also namely with the pulse output signals of STDP plasticity effect.In the present embodiment, the output of K1601 is also with the network that R1606, R1607, R1608, D1605, C1603 are formed, and for simulating the cynapse depression effect of journey in short-term further, its principle of work can with reference to the analysis of Figure 12 circuit.
The analogy method of " cynapse of prefabricated property " of the present invention comprising: (1), between the cephacoria input end and caudacoria output terminal of cynapse preset a cynapse transmission channels, but the work of this cynapse transmission channels of closedown that begins of work; (2), detect the spike signal of presynaptic membrane input end and caudacoria output terminal; (3), between cephacoria input end and caudacoria output terminal, there is the spike signal meeting STDP characteristic, then computing is carried out to the appearance situation of this signal, (comprising integration or counting); (4), when the satisfied setting rule of situation that the spike signal meeting STDP characteristic occurs, then start the work of default cynapse transmission channels, the cephacoria input end of cynapse and caudacoria output terminal are set up there is the transmission channels that usefulness is transmitted in cynapse.
Described default cynapse transmission channels, a cynapse transmission channels with cynapse transmission characteristic, one or more in the synaptic plasticity that these cynapse transmission characteristics comprise synaptic facilitation, cynapse suppression, the tetanic rear enhancing of cynapse, cynapse Long-term plasticity particularly have STDP characteristic, these transmission characteristics.In other words, default cynapse transmission channels of the present invention, it has been exactly in fact the cynapse analogue means with transmission characteristic of a prior art, it can be aforesaid " fixing cynapse " device, also can be aforesaid " modifiable synapse " device, or the cynapse analogue means of other prior aries.
Described setting rule is the one in following two kinds: 1, only counting meets the number of times of the signal of the LTP feature in STDP characteristic, if within the time of one section of setting, occurs that the counting of LTP reaches setting value, then starts the work of default cynapse transmission channels; Or 2, do not set gate time, counting meets the number of times of the signal of LTP characteristic sum LTD feature simultaneously, wherein occurs that the signal number of times of LTP feature does addition, occurs that the signal number of times of LTD feature does subtraction; If tale reaches setting value, then start the work of default cynapse transmission channels.
Further improvement, the analogy method of " cynapse of prefabricated property " of the present invention, also comprises:
(5), after the work starting the cynapse transmission channels preset, continue the spike signal detecting presynaptic membrane input end and caudacoria output terminal; (6), between cephacoria input end and caudacoria output terminal, there is the spike signal meeting STDP characteristic, then computing is carried out to the appearance situation of this signal; When the spike signal meeting STDP characteristic occurs that situation meets the second setting rule, then close the work of default cynapse transmission channels, cut off the cephacoria input end of cynapse and the transmission channels of caudacoria output terminal.
The second described setting rule is the one in following two kinds: 1, only counting meets the number of times of the signal of LTD feature, if within the time of one section of setting, occurs that the counting of LTD reaches setting value, then closes the work of default cynapse transmission channels; 2, counting meets the number of times of the signal of LTP characteristic sum LTD feature simultaneously, wherein occurs that the signal number of times of LTP feature does addition, occurs that the signal number of times of LTD feature does subtraction; If tale is lower than setting value, then close the work of cynapse transmission channels.Obviously in rear a kind of this setting rule, close the setting value of the tale of cynapse transmission channels, the setting value of the tale starting cynapse transmission channels must be less than.
As Figure 17.The analogue means of " cynapse of prefabricated property " of the present invention, includes cephacoria input end, caudacoria output terminal and has the cynapse transmission channels that usefulness is transmitted in cynapse; It is characterized in that: also with a STDP chronicle recognition circuit, for identifying the spike signal between cephacoria input end and caudacoria input end with STDP feature; The output of STDP chronicle recognition circuit is connected to a STDP integrating circuit, and the output of STDP integrating circuit is connected to a thresholding control circuit; The output terminal of thresholding control circuit is connected to a channel switching circuit; Channel switching circuit is for controlling the break-make of cynapse transmission channels.
Its principle of work and process are: during work, and STDP integrating circuit, according to setting rule, is accumulated the signal that the institute of STDP chronicle recognition circuit within a certain period of time identifies output, and is converted into its output valve outputed signal; Thresholding control circuit, for detecting the output valve of the output signal of integrating circuit, reaches when the output valve of integrating circuit or higher than setting thresholding 1, then produces and export control signal; And start connection cynapse transmission channels by channel switching circuit, set up between cephacoria input end and caudacoria output terminal and there is the cynapse transmission channels that usefulness is transmitted in cynapse.
Its course of work further comprises: after channel switching circuit starts connection cynapse transmission channels, thresholding control circuit continues the output valve of the output signal detecting integrating circuit, when the output valve of integrating circuit is lower than setting thresholding 2, then closes and export control signal; And close cynapse transmission channels by channel switching circuit.Also the cynapse transmission channels between cephacoria input end and caudacoria output terminal is namely cut off.Obviously, setting thresholding 2 needs to be less than setting thresholding 1, more reasonably arranges, generally can be equivalent to 65% to 85% of the value setting thresholding 1 as one.
Same, described default cynapse transmission channels is a cynapse transmission channels with cynapse transmission characteristic.
The setting rule of described integrating circuit is: the output valve improving output signal when there is the signal of LTP feature, reduces the output valve of output signal when there is the signal of LTD feature.(relative to analogy method, analogue means only adopts wherein a kind of mode closer to brain work characteristics to the operation rule of STDP).
The analogy method of " cynapse of prefabricated property " of the present invention and device, transmit compared with the plastic cynapse analogue technique of STDP with the cynapse that has of prior art, although all have the technology meeting STDP characteristic spike signal between detection and Identification presynaptic caudacoria each other, its process made is different.Existing STDP cynapse analogue technique, initially there is cynapse transmit usefulness, again according to detecting the spike signal meeting STDP characteristic identified, adjust cynapse and transmit usefulness: if there is LTP characteristic signal, then promote cynapse and transmit usefulness, if there is LTD characteristic signal, then reduce cynapse and transmit usefulness, realize cynapse with this and transmit plasticity.From brain structure, this class cynapse is present between the neuron in the regions such as amygdaloid body, corpus straitum, hippocampus more, for realizing journey in short-term and the long-term memory of declarative information.
And " cynapse of prefabricated property " of the present invention, itself just include one and possess the cynapse transmission channels that usefulness is transmitted in cynapse, also the cynapse analogue means of a prior art namely itself is included, be on the basis of the cynapse analogue means of prior art, add a technology how identifying, judge, start or close cynapse analogue means.Its cynapse transmission channels when working original state is closed, so do not possess cynapse at the beginning to transmit usefulness, only has the spike signal repeatedly occurring the LTP feature met in STDP characteristic between cephacoria input end and caudacoria output terminal, and accumulation reaches a certain degree, just can start the work of connecting cynapse transmission channels by channel switching circuit, make " cynapse of prefabricated property " between cephacoria input end and caudacoria output terminal, possess effective cynapse and transmit usefulness.What its was simulated is a kind of cynapse be present between the neuron on the regions such as cortex, do not possess cynapse when this kind of cynapse starts to be formed and transmit usefulness, need after the neuronic collaborative stimulation repeatedly for a long time in front and back, just be activated into effective cynapse, and once activate, it is just more stable that it transmits usefulness.This kind of cynapse is generally present in the regions such as cortex, for realizing the long-term memory of declarative information.
Spike signal, as long as or have and to be made by its cynapse transmission presynaptic neuron be activated to produce action potential (actual this is also produce LTP feature), then the numerical value of STDP integrating circuit is just more and more higher until reach maximal value, and also namely cynapse transmission usefulness can be more and more tetanic to maximal value.Only have and do not occur LTP feature for a long time, and repeatedly occur the inhibition signal of LTD feature, the numerical value of integrating circuit just can be made slowly to reduce, final lower than threshold value, thresholding control circuit cuts out output signal, closes cynapse transmission channels, cynapse was lost efficacy again by channel switching circuit.Even information preserves as long-term memory for simulating cerebral neuron by this duty, if these information do not use for a long time, and having obscuring for a long time (competition) of other inhibition information, then long-term memory also can slowly pass into silence.
" cynapse of prefabricated property " of the present invention analogue means, STDP chronicle recognition circuit belongs to prior art, and how to LTP and the LTD characteristic signal that STDP chronicle recognition circuit exports, computing is carried out for STDP integrating circuit, generally the microprocessors such as MCU can be adopted, according to described setting rule, count operation is carried out to LTP characteristic signal and LTD characteristic signal, again count value is compared with setting value, export control signal according to comparative result, the break-make of the channel control circuit of cynapse transmission channels is controlled.That is, same single-chip microcomputer can be adopted to complete the function of chronicle recognition circuit, STDP integrating circuit and thresholding control circuit simultaneously, so hardware components is very simple, not draw here.And for the program of single-chip microcomputer, person skilled in art, principle disclosed according to the present invention and operation rule, can both write these programs easily.
The present invention is also with reference to the analogue technique of " modifiable synapse " of aforementioned Figure 15, and open another kind of employing accumulate device realizes the technology of " cynapse of prefabricated property " analogue means of STDP integrating circuit.Its circuit block diagram is as Figure 18.STDP integrating circuit adopts charging circuit, discharge circuit and an accumulate device to realize, and the LTP output terminal of STDP chronicle recognition circuit and LTD output terminal are connected on accumulate device respectively by charging circuit and discharge circuit; The output terminal of accumulate device is connected to the input end of thresholding control circuit.During work, STDP chronicle recognition circuit LTP output terminal when identifying LTP feature occurs that pulse exports, and is charged to accumulate device by charge circuit; STDP chronicle recognition circuit LTD output terminal when identifying LTD feature occurs that negative pulse exports, and by discharge loop, accumulate device is discharged; Change is there is in the voltage of accumulate device with STDP effect; The voltage of thresholding control circuit to accumulate device detects, and when accumulate device output voltage meets or exceeds setting threshold value 1, thresholding control circuit exports control signal, and starts connection cynapse transmission channels by channel switching circuit.
Its course of work further comprises: after channel switching circuit starts connection cynapse transmission channels, thresholding control circuit continues the output voltage detecting accumulate device, when the output voltage of accumulate device is lower than setting thresholding 2, then closes output control signal; And close cynapse transmission channels by channel switching circuit.Also the cynapse transmission channels between cephacoria input end and caudacoria output terminal is namely cut off.Obviously, set thresholding 2 and be less than setting thresholding 1.More reasonably arrange as one, setting thresholding 2 can be equivalent to 65% to 85% of the value setting thresholding 1.
Figure 19 is the circuit theory diagrams of embodiment.The signal of cephacoria input end and caudacoria output terminal is inputted by I/O1 and I/O2 of MCU1801, the LTP signal and the LTD signal that pass through identification are exported by I/O3 and I/O4 respectively, (be that positive output and negative output are effective respectively, and output pulse width has STDP effect, principle of work with reference to Figure 16), and by R1801 and R1802 electric capacity C1801 charged and generate electricity, the voltage on C1801 is changed along with STDP effect.The magnitude of voltage of C1801 provides R1803 to deliver to I/O5 more again, voltage compare is carried out according to setting rule by the voltage comparator circuit inside MCU1801, (also can carry out voltage compare again after first analog to digital conversion), its comparative result exports from I/O6, go to control the break-make as the analog switch K1801 of channel control circuit, thus realize controlling the break-make of cynapse transmission channels.
Cynapse transmission channels has cynapse transmission characteristic, can reference and adopt existing cynapse analogue technique, also can adopt the analogue technique of the present invention's aforesaid " solidification cynapse ".Figure 20 is a kind of embodiment of " cynapse of prefabricated property " analogue means.The wherein K1801 of the MCU1801 of MCU2001 also i.e. Figure 19, analog switch K2001 also i.e. Figure 19, its cynapse transmission channels is made up of C2002, R2004, R2005, R2006 and D2003 etc., and its principle of work can with reference to the explanation of Figure 12.
For STDP chronicle recognition circuit, prior art generally adopts MCU circuit, and according to the STDP family curve of Figure 13, employing program identifies and exports.The present invention discloses a kind of technology adopting common digital gate circuit to identify to realize STDP and export here.Circuit is as Figure 21.From the action potential spike potential signal of cephacoria input end and caudacoria output terminal respectively through R2101, R2102, dividing potential drop is carried out with R2103, R2104, the spike potential making it input must reach enough magnitudes of voltage could effectively trigger not circuit F2101, F2103, avoid interference, input spike potential is once effectively trigger F2101, F2103, just the monostable circuit of F2102 or F2104 composition can be triggered, export a rect.p., (pulse width is approximately 5-10 times of action potential pulsewidth); F2105-F2108 forms interlock circuit, makes F2107 and F2108 can only have an output, (F2102 or F2104 first triggers and export rect.p., F2107 or F2108 just exports positive level, and pin another side be output low level); When F2102 and F2104 exports rect.p. simultaneously, it is just that AND circuit F2112 exports, its effective width exported that is to say the time-interleaving part of the rect.p. that F2102 and F2104 exports separately, obviously relevant to the sequential of the spike potential signal of cephacoria input end and caudacoria output terminal.The output of F2112 and interlock circuit, is connected to F2109 and F2110.When the spike potential signal of cephacoria input end is early than caudacoria output terminal, when also namely there is LTP characteristic signal, F2109 exports positive level as LTP output terminal, and its pulse width negative exported is about the mistiming of front and back spike potential, (as Figure 13); When the spike potential signal of cephacoria input end is slower than caudacoria output terminal, when also namely there is LTD characteristic signal, F2110 exports positive level as LTD output terminal, and its pulse width negative exported is about mistiming of front and back spike potential, and through F2111 anti-phase after export negative level as LTD output terminal.Diode D2101 and D2102 plays unidirectional output, that is to say in aforementioned Figure 16, Figure 19 and Figure 20, the diode that LTP output terminal and LTD output terminal connect.This circuit adopts common and cheap digital gate circuit, realizes the detection to STDP spike potential and output dexterously, although its output accuracy is not very high, but has the advantage of very low cost, and just can work without the need to relying on manual program.
The present invention also discloses the simulative neural network that two kinds have different working characteristic.
As Figure 22.A kind ofly neuron analogue means of the present invention is adopted to form, there is the neurn simulation network of imformation memory and reflection function, comprise at least one neuron matrix unit, each neuron matrix unit comprises multiple (multiple i.e. two or more) presynaptic neuron analogue means (as the A1 to An in Figure 22 and B1 to Bn), with multiple postsynaptic neuron analogue means (C1 to Cn in Figure 22), between the aixs cylinder output terminal of presynaptic neuron analogue means and the dendron input end of postsynaptic neuron analogue means, employing has cynapse transmission usefulness STDP plastic cynapse analogue means PY and connects, (these are similar with Fig. 3), it is characterized in that: " the attentional manipulation end " of each neuron analogue means links together, by a road, only there is low and high level two states, control end AT for the memory/reflection two kinds of duties switching neuron analogue means controls.As the simplest a kind of mode, described control signal is provided by a duty change-over switch.
Neurn simulation network of the present invention, can by neuron analogue means and between connect there is the plastic cynapse analogue means of cynapse STDP, realize the input memory of information and reflection.Its principle of work is similar with the technology of existing neurn simulation network: when neuron is in memory (study) duty, if some presynaptic neuron (A1 to An and B1 to Bn, be equivalent to input neuron) generation action potential is activated due to the stimulation of input signal, then can make some postsynaptic neuron (C1 to Cn by cynapse integration, be equivalent to output neuron) be also activated and output reflection signal, if former and later two information of input have spatial correlation (can realize cynapse integrate) and temporal associativity (successively and followed by appearance, also be so-called Time Dependent), the cynapse PY existed between them then can be made to occur STDP synaptic plasticity, the signal reflex path of neural network is changed by the change of cynapse transmission usefulness, thus new reflection channel is set up between two information.And when neuron is in reflective operation state, if input last information at presynaptic neuron, reflection channel then by having set up, just postsynaptic neuron can be made to reflect and to export a rear information, realize the Memory acquisition (note) of two information and again obtain the function of (recalling).
But, compared with the neurn simulation network of prior art (as Fig. 3), neurn simulation network of the present invention, no matter be made up of how many neurons, also how huge the neuron matrix unit quantity no matter formed have, all neuronic attentional manipulation ends need link together, be controlled by the control signal AT of same road memory/reflective operation state by all the same.When control signal AT is placed in " reflection " state, neurn simulation network carries out reflection process (also namely simulating thinking and the reaction of brain) in the mode of reflection to input information, only to the unidirectional output of aixs cylinder during neuron snapping motion current potential, cynapse analogue means can not produce transmission usefulness STDP plasticity owing to lacking the spike signal of caudacoria, and the processing procedure of information can not cause interference to the recall info (experience, rule, custom etc.) originally formed; When control signal AT is placed in " memory " state, the neuron of neurn simulation network processes (learning and memory of also namely simulating brain) input information with the working method of memory, both to axonal transport during neuron snapping motion current potential, also to dendron back transfer, cynapse analogue means is enable to produce STDP plasticity.In other words, neurn simulation network of the present invention is when carrying out imformation memory, the pulse of spike accurately required for its generation cynapse STDP plasticity and sequential thereof, (as the brain) that produced by that neuron activated by excitement oneself, and without the need to producing additionally by other external circuits.So, neurn simulation network of the present invention is without the need to arranging many clock control ends, more accurately control without the need to the sequential of manual program to the spike of clock control end of being unified by department of computer science, so, the system architecture of neurn simulation network of the present invention is simpler, also carrying out work without the need to depending on department of computer science's manual program of unifying, the work of the brain of natural evolution can be simulated more truly.
The present invention also openly another kind there is short-term memory and long-term memory and short-term memory can be converted into the simulative neural network of long-term memory in any case.As Figure 23, described simulative neural network comprises the first input neuron group IN1, the first output neuron group OUT1, the second input neuron group IN2, the second output neuron group OUT2, the first intrerneuron group M1 and the second intrerneuron group M2.Here the neuron analogue means that neuron can adopt the function of prior art suitable, also can adopt neuron analogue means disclosed by the invention.Each neuron pool by least 2, be generally that the neuron of One's name is legion forms, in order to accompanying drawing is clear, each neuron pool of Figure 23 only draws two neurons.And, it should be noted that: in order to accompanying drawing structure is clear and be convenient to principle of work is described, Figure 23 is divided into IN1 and IN2 of different directions input neuron group, output neuron group is divided into OUT1 and OUT2, but at work, IN1 and IN2, and between OUT1 and OUT2, difference not in job specification, they all belong to the same input neuron and output neuron, and just position is different.(on brain, they are all the neurons of the information reflex path be present on cortex).
The feature of this neurn simulation network is: the aixs cylinder output terminal of the first input neuron group IN1 is connected with adopting " solidification cynapse " SY1 between the dendron input end of the first output neuron group, adopts " solidification cynapse " SY2 to be connected between the aixs cylinder output terminal of the second input neuron group IN2 and the dendron input end of the second output neuron group OUT2; (neuron being similar to the stable information input and output passage formed in brain connects);
The aixs cylinder output terminal of the first input neuron group IN1 and the second input neuron group IN2, and between the dendron input end of the first intrerneuron group M1, adopt " modifiable synapse " PY to connect; (be similar to cerebral cortical neuron by the projection to hippocampal dentate granular cell of the path of perforating fiber, and adopt between the two there is the plastic cynapse of STDP connect);
The aixs cylinder output terminal of the first intrerneuron group M1, and between the dendron input end of the second intrerneuron group M2, adopt " modifiable synapse " PY to connect; (be similar to inside hippocampus, the mossy fiber that granular cell sends, and the connection with the plastic cynapse of STDP between hippocampus CA3 cones);
The aixs cylinder output terminal of the second intrerneuron group M2, and between the dendron input end of the first output neuron group OUT1 and the second output neuron group OUT2, adopt " modifiable synapse " PY to connect; (Schaffer that the CA3 cones being similar to hippocampus sends is collateral, and has the plastic Synaptic junction of STDP between CA1 cones);
Between the aixs cylinder output terminal of the first input neuron group IN1 and the dendron input end of the second output neuron group OUT2, between the aixs cylinder output terminal also having the second input neuron group IN2 and the dendron input end of the first output neuron group OUT1, " cynapse of prefabricated property " GY2 with GY1 is adopted to be connected; (the similar connection that there is prefabricated property cynapse between cerebral cortex epineural unit, and can activate in any case and become effective cynapse, thus realize the conversion of short-term memory to long-term memory).
The principle of work (in fact that is to say that brain forms the Forming Mechanism of short-term memory such as hippocampal long-term memory to two information) that this neurn simulation network forms short-term memory is: the first input neuron group IN1 and the first output neuron group OUT1, and the second input neuron group IN2 and the second output neuron group OUT2, by the connection of effective curability cynapse strong inside them, output neuron group energy is made enough to carry out reflection process to the input signal of input neuron group, form the reactive reflex path of two barss, (being similar to the stable information reflection channel formed in brain).Meanwhile, these two reactive signal reflex paths produce annexation by two projection tracts and two intrerneuron groups again simultaneously, form the intrerneuron process path of signal.Wherein the aixs cylinder output terminal of input neuron group IN1 and IN2 projects the dendron input end of the first intrerneuron group M1 all simultaneously, form the input projection tract of intrerneuron process path, the aixs cylinder output terminal of the second intrerneuron group M2 projects the dendron input end of the first output neuron group OUT1 and the second output neuron group OUT2, forms the output projection tract of intrerneuron process path.
When on input neuron group time of occurrence relevant (successively occur and followed by, i.e. Time Dependent) two signals (the first input signal and the second input signal), also namely activating successively appears in some neuron of input neuron group IN1 and IN2, form action potential when exporting spike, except being produced two output signals (the first output signal and the second output signal) respectively on output neuron group by original reactive reflex path, also make some neuron in the first intrerneuron group M1 also occur activating formation action potential by cynapse integration simultaneously, these neuronic activation of first intrerneuron group M1 export, and can be integrated equally again make some neuron inside the second intrerneuron group M2 also activate output spike by cynapse, and the aixs cylinder output of the second intrerneuron group M2 is the dendron input end projecting output neuron group OUT1 and OUT2, this spike signal that M2 aixs cylinder is exported, understands with the state of activation of output neuron group OUT1 and OUT2, also namely the signal of two input and output neuron signal reflex paths inputs with reflective condition closely related in time, if at this moment neuron is in " memory " duty, then neuronic activation can produce the back transfer of action potential, also namely its dendron input end there will be the spike of action potential back transfer, make the cynapse on intrerneuron process path can produce STDP synaptic plasticity, (synaptic plasticity wherein between M1 and M2 and between M2 and OUT1 and OUT2 is more remarkable), some cynapse is transmitted usefulness and is enhanced due to LTP effect, and some cynapse is transmitted usefulness and is inhibited and weakens due to LTD.If same situation (namely occurring two signals that room and time is relevant on input/output signal reflex path) repeatedly occurs within a certain period of time, then these plasticity effects can be strengthened gradually, make between input/output signal reflex path and intrerneuron process path, form out the middle reflex path that can reflect two signals gradually, (concrete middle reflex path is: input neuron group IN1 or IN2 → → intrerneuron group M1 → → intrerneuron group M2 → → output neuron group OUT1 or OUT2), also namely between the neuron corresponding to two signals, relevance reflective relation is established, this is the process of the acquisition (note) of memory.
Afterwards, if again input first signal (the first input signal) having occurred setting up in two correlation signals of memory relation input neuron group, also the neuron (can be or one group of neuron) namely corresponding to the first input signal in input neuron group occurs activating due to outside input, then this signal is except directly reflecting the first output signal output neuron group by original reflection channel, also can make reflex response by the middle reflex path set up, the neuron (can be one or one group equally) corresponding to second signal in output neuron group is made to occur activating, thus make output neuron group output the second output signal, the reflection realizing " the first input signal → → the second output signal " exports, this is the process obtaining (recalling) again of memory.This memory is realized by the Long-term plasticity of cynapse, belongs to long-term memory, be similar to the declarative memory of hippocampus in brain, or striatal procedural memory, or the emotionality memory of amygdaloid nucleus.
The principle of work that this neurn simulation network forms long-term memory is: on the above-mentioned basis having formed short-term memory (long-term memory), if occur that on input/output signal reflex path the situation continued of two coherent signals repeatedly occurs, this may be on input neuron group IN1 and IN2, continue repeatedly successively input appearance first input signal and the second input signal reflect output first and output signal and second to output signal, (being similar to the process of continue studying consolidating memory after forming memory), also may be on input neuron group, repeatedly input the first input signal and reflect output second by middle reflex path to output signal, (be similar to and recalling or in thinking, realizing the further process strengthening consolidating to original memory), and, if at this moment neuron is in learning state, so, the neuron that the signal reflex path of information input and output successively activates, the spike signal of action potential when it activates, constantly can form relevance to " cynapse of prefabricated property " i.e. existing between them GY1 and GY2 to stimulate, wherein a part " cynapse of prefabricated property " finally can activate and becomes effective cynapse owing to being repeatedly subject to the just stimulation (LTP effect) of STDP, another part is repeatedly subject to negative stimulation (LTD effect) and is suppressed by continuation, (its principle of work sees above describing of face relevant " cynapse of prefabricated property " part), thus between input neuron group and output neuron group, the reflex path of directly " the first input signal → → the second output signal " is set up by activating " cynapse of prefabricated property ".
At this moment, if again input input neuron group and occurred the first input signal, then neural network is without the need to the indirectly middle reflex path again by being set up by intrerneuron group M1 and M2, and directly by between input neuron group and output neuron group by the direct reflex path activating " cynapse of prefabricated property " and set up, reflect output second to output signal, the memory namely achieving " the first input signal → → the second output signal " obtains the process of (recalling) again.The association needing time more grown or more secondary due to this direct reflex path stimulates and could set up, once set up, can be more stable, and be not easy, again by impact and the destruction of other input signals, so can be described as long-term memory, or to be called permanent memory.In the brain, this memory is usually recorded in the regions such as the cortex of brain, so also can be described as cortex memory.
And once set up long-term memory by activating " cynapse of prefabricated property ", the then middle reflection channel of short-term memory set up of original " modifiable synapse " by intrerneuron, then can due to the foundation of " cynapse of prefabricated property " this direct reflection channel, and the output signal of middle reflection channel is always occurred slowly than the output signal of direct reflection channel, also namely the spike of presynaptic membrane is slower than the spike of caudacoria, thus the cynapse producing STDP suppresses, final this indirectly in the middle of reflection channel lose effect, and discharge, for setting up the short-term memory of new information.
Simulative neural network of the present invention, can realize between different information IO channel, by intrerneuron and between connect " modifiable synapse " realize short-term memory, and, if this short-term memory is repeatedly started or uses, then can activate and be preset at " cynapse of prefabricated property " between information IO channel, thus between two information IO channel, set up more directly more stable cynapse transmission channels, realize long-term memory.
It should be noted that, in circuit diagram disclosed in present specification, generally only draw and mark the element relevant with principle of work, and the startup of power supply supply, power supply voltage stabilizing, anti-interference, single-chip microcomputer to be resetted etc. peripheral interlock circuit, belong to the routine techniques of electronic applications, then in order to accompanying drawing succinctly and not marks.
In various analogue technique of the present invention, except the principle design of circuit, also has the setting of some major parameters.In the real work of brain, the time span of various work time-histories is large especially, such as: the pulse width of action potential is from several milliseconds to a few tens of milliseconds; The integration process of membrane potential of neurons greatly about a few tens of milliseconds to millisecond up to a hundred; Time in possible a few second of modularity cynapse generation modulating action is to dozens of minutes; Hippocampal neuron can from a few minutes by several hours to several days for the time of short term plasticity during Memory Activities and Long-term plasticity; And the cynapse of " prefabricated property " is activated the cynapse developing into " validity " by the neuron of cortex, or newly grow an effective cynapse, then need a lot of sky even time of several years.For the work of neurn simulation network, neurn simulation network is particularly adopted to test and demonstrate the work of neuron activity, same time span is adopted to be unpractical, so conveniently we can only according to need of work, according to the length relation of various work time-histories, set the time value of each work time-histories separately.In addition, potential change in neural network and signal amplitude are the magnitude of voltage of several millivolts to tens millivolts mostly, if mimic channel adopts this order of magnitude to carry out work, then both being difficult to arrange also easily works abnormal, so we also can only set the magnitude of voltage of various signal and change separately by electromagnetic interference (EMI).
As a kind of example, below provide a kind of numerical value of each work time-histories and various signal voltage, this example is suitable for neuron analogue means, cynapse analogue means to make experiment equipment, for constructing neurn simulation network, the more working mechanism of brain is studied and disclosed to simulation and demonstration neuron and cynapse, to active situation such as the reflection of signal and memories, and with this.
Analogue means operating voltage: DC12 lies prostrate.
Action potential pulse width: 10 milliseconds.
Action potential peak value of pulse: >=11 volts.
First threshold trigger circuit setting activation threshold value: 3 volts.
Second Threshold trigger circuit setting activation threshold value: 3.7 volts.
The time delay of the signal delay circuit of first threshold trigger circuit input end: about 20 milliseconds.
The signal integration trigger condition of film integrating circuit: the exciting signal of input makes the output voltage of film integrating circuit reach first threshold trigger circuit setting activation threshold value in 100 milliseconds.
Refractory period after action potential triggers: about 100 milliseconds.
Modularity signal (comprising enhancement and the inhibition) term of validity: about 10 seconds.
Postsynaptic membrane output signal master output: 8 volts, (during without synaptic plasticity).
Postsynaptic membrane output signal amplitude peak: 10 volts, (when synaptic plasticity LTP is maximum).
Postsynaptic membrane output signal minimum radius: 6 volts, (when synaptic plasticity LDP is maximum).
The long term synaptic plasticity term of validity of " modifiable synapse ": about 5 minutes.
The setting threshold value 1:8 of the thresholding control circuit of " cynapse of prefabricated property " lies prostrate.
The setting threshold value 2:5 of the thresholding control circuit of " cynapse of prefabricated property " lies prostrate.
" prefabricated property " cynapse activates a kind of condition for " validity " cynapse: in 10 minutes, occur the front and back neuron spike signal meeting the LTP effect of STDP rule for 10 times.
According to above numerical value, then according to the theory of electricity and formula, just can set and calculate the concrete numerical value of the Resistor-Capacitor Unit of each RC circuit and bleeder circuit.Certain above-mentioned numerical value can adjust according to the actual needs of lecture experiment.
In addition, applicant is according to the analysis to cerebral neuron and cynapse working mechanism, and the simulated experiment about memory using analogue means of the present invention to carry out, and infers the essence about brain memory function and working mechanism.Below this is simply introduced.These contents may not have direct relation with the realization of the concrete technology of the present invention, but the working mechanism of the present invention, particularly " modifiable synapse " and " cynapse of prefabricated property " and the not same-action in memory is formed thereof can be described better.
One, at present about the defect that the plasticity theory of learning and memory mechanism exists.
1.1 about the mechanism of learning and memory, current widely accepted prevailing be " synaptic plasticity " theory coming from " Hebb hypothesis ".This theory thinks that learning and memory is that the stimulation of external information causes the structure of cynapse between neuron or formation to change, cynapse is caused to transmit the change of effect, i.e. long term potentia ̄tion (LTP) and long-term depression (LTD), this change encodes the information of acquisition, defines memory.This theory obtains the support of abundant experimental results in decades, and be especially formation and the maintenance of explaining short-term memory, this theory obtains the support of anatomy and a large amount of observe phenomenas biologically.And cynapse STDP plasticity is supplementing further " Hebb hypothesis ", it disclose synaptic plasticity and relation that after the presynaptic, neuron produces excited sequential: when presynaptic neuron excited prior to rear neuronal excitation time, produce long term potentia ̄tion (LTP), when rear neuronal excitation is prior to front neuronal excitation, produce long-term depression (LTD), if the front and back neuronic excited difference time is comparatively large, does not then produce plasticity and modify.
But there are some defects that cannot overcome in this theory, cannot explain some memory phenomenons, and this comprises following content.
1.2 synaptic plasticity, the formation comprising long term potentia ̄tion (LTP) and long-term depression (LTD) often only needs a few minutes to several hours, but some memory often need several days, a few week, so that in the past few years formed and consolidate, the same synaptic plasticity, the same synaptic structure, why can cause so different memory difference? the effective period of synaptic plasticity only has several days to a few week, maximum some months, also cannot explain the phenomenon that a lot of memory can be preserved down to decades for several years for a long time.Academic circles at present is often interpreted as cynapse simply through repeatedly stimulating, and the permanent vestige of Synaptic formation realizes long-term memory, but how to form permanent vestige for synaptic plasticity, and what feature the cynapse formed after permanent vestige has, then cannot illustrate.
The 1.3 synaptic plasticity theories that memory is formed, " note " and " recalling " occurs on same neuron, the input stimulus neuron of information causes producing synaptic plasticity, formed " note ", when again stimulating this neuron, existence due to synaptic plasticity causes neuronic activation action, and produces " recalling ".But we know, declarative memory is in forming process, the memory of different phase is the diverse location being kept at brain: in the initial period of memory, particularly in so-called " working memory " stage, memory occurs on cerebral cortex, if but remember consciously, then the formation of long-term memory depends on the participation of hippocampus, (main foundation is: 1, hippocampus is extractd cannot form long-term memory, 2, hippocampus excision can cause the antidromicity of time laminarity to be lost memory), and when the time long enough of long-term memory or long-term memory enough firmly time, memory seems again to be again kept on cortex, (hippocampus is extractd does not affect these memories).And these change and obviously also do not need you deliberately to go to transform, but unconsciously with regard to physical alterations.These features of memory, the body extracing people can obviously show, the typical example as famous patient H amnesia or hippocampus.Existing theory is obviously difficult to the problem interpreted these phenomena, especially how short-term memory is converted into long-term memory.
The core of the problems referred to above is exactly the problem of the Forming Mechanism of long-term memory.That is, synaptic plasticity theory can explain the Forming Mechanism of short-term memory well, also obtains the support of great many of experiments, but cannot explain the formation of long-term memory.
1.4 for the Forming Mechanism of long-term memory, and some scholar proposes some new theories at present.Such as " genomic theory ", thinks that long-term memory is the modification relying on protein on gene chromosome, even relies on the modification of DNA itself to carry out.But the experiment that this theory obtains is supported less.Particularly, this theory is only for long-term memory, and invalid to short-term memory, and like this, brain just needs employing two kinds of diverse mechanism to carry out Memory Activities, and this does not meet the feature of biological evolution very much.And, due to be employing two kinds of diverse mechanism to form short-term memory and long-term memory, so we formed and change such two kinds memory time will inevitably there is sensorial difference, even cause the confusion of some tomographies and intersection.But we never feel such difference, short-term memory and long-term memory do not have obvious boundary at all.
Two, the essence of memory, and the formation of memory.
Applicant analyzes the existing various well known phenomenon about human brain memory, the essence remembered and Forming Mechanism are set forth, this theory can be compatible and explain the formation of short-term memory and long-term memory uniformly, and the anatomical features of brain can be met on microcosmic, macroscopically also can explain the various phenomenons of brain about memory very well.
2.1 first, and from memory macro phenomenon, imformation memory comprises " note " and " recalling " two processes, by information input and preserve process be " note ", the process that information is output under certain condition is " recalling ".From the input/output procedure of information, information is not isolated to exist, but there is correlative connection with other information.Conform to the principle of simplicity unilaterally, what does is the extraneous information such as sound, image, sensation? how? how to change? all must there is object, or with other information-related relations.Even if from declarative memory.Such as, pi equals 3.14, and how Zhang San looks, and I eats face for lunch, and newton is physicist, and you went Europe tourism with wife last year, etc., these related informations link and remember by the input memory of information in the lump.When information is output conversely, circular constant → π → equal → 3.14, Zhang San → look → how, I → lunch → eat → face, newton → be → physicist, you → last year → with → wife → go → Europe → tourism, be also that a certain information or the input of some information of these related informations makes another information associated be output, complete " recalling " of information.
The most isolated situation, allows you remember a unrelated noun " cloud " now, so, " cloud " in fact also in front of face " allow now you remember a noun " these information are associated.After you complete the work of note, if ask you " what yesterday noon eats ", you scarcely can just now remember that " cloud " answers out, and only ask you " what the noun that you remembered just now is ", you just can export " cloud " this information because of the input of these related informations.
Each information is also not only organized information-related with another or another, but all relevant with other bulk informations, different related informations defines different a certain customizing messages, and, only have when all related informations all input, could determine and export limited customizing messages.Such as, my breakfast drink milk, I eats face for lunch, and he eats chicken for lunch, and he is named as Li Si, my breakfast a people eat, I runs into Li Si for lunch, in these information, if when only having input information " I " or " he ", cannot limit and export other information completely, because have relevant information all too many with " I " or " he "; And when inputting " I " and " lunch ", scope obviously reduces a lot, also still cannot clearly limit, because may be " I has face for lunch " also may be " I runs into Li Si for lunch "; When inputting " I ", " lunch ", " eating " successively, " face " this information uniquely can be limited and is exported.Same, when inputting " I ", " breakfast ", " drinking " successively, then " milk " this information also uniquely can be limited and is exported.
So from macroscopic view, the memory of information is not isolated existence, but need to follow the link of other related informations to exist.A certain information is set up fixing incidence relation with other or one group of information, and this is " note "; And when related information inputs, setting up relevant a certain information is exported accordingly, this is " recalling "; After the process of " note " completes, correctly can complete the process of " recalling " during input related information, just complete the function of " memory ".
So 2.2, we are from macroscopically so stating the memory of information: " exist the information bank of information at one; no matter its quantity of information has much also how complicated regardless of its relation; if information A and information B sets up so a kind of relation; make as input information A; information bank has and only has (uniquely) output information B, so, we just can think that information A and information B achieves memory relation.I.e. information A → information B ".
Here information A is input information, and A may be an information, also may be one group of information, (such as " I ", " lunch ", " eating "), and information B is output information (such as " face ").And information B is once be output, the input information of next process may be become again, jointly cause the output of another information C of next process with A.That is: information A+ information B → information C.The like, so that the chain type forming an information exports.
2.3 neuromechanisms getting back to brain.From the Nomenclature Composition and Structure of Complexes of brain, brain is made up of neuron, and so the thinking of brain is necessarily realized by the activity of neuron and cynapse connected to each other thereof with memory.And we know, each neuron not only forms synaptic connections with another neuron, but form complicated synaptic connections with other a large amount of neurons, neuron can by cynapse transmit activate and trigger, often neither only be triggered by an action of a certain neuronic some cynapses, but need dependence cynapse integration to realize, by a series of transmission action (time of cynapse is integrated) of certain cynapse, or action while of very multisynaptic (spatial integration of cynapse), just can cause a neuronic effective activation.(in corticocerebral information processing, the spatial integration of cynapse seems more main).In other words, for carrying out the senior neururgic cerebral nerve metanetwork such as thinking and memory, the activation action of a certain neuron A, transmitted by cynapse and make another neuron B be triggered and activate, often be not only and realized by A, B a certain isolated cynapse transmission between the two, and need the acting in conjunction of other related Neurons depended on around neuron B often, integrated by cynapse and to realize.This so-called acting in conjunction, that is to say that the related Neurons around neuron B was once activated within a period of time not long ago that neuron A is activated, certain stimulation is transmitted to neuron B by the cynapse between they and neuron B, but these stimulations are still not enough to trigger neuron B and produce action potential formation activation (integrate and still do not reach activation threshold value), and can only as a kind of excited, make neuron B more easily by follow-up other stimulate trigger.
But, other neuronic effects just basis and prerequisite around neuron B, the activation of neuron A is only key, just because of the basis of the cynapse integration that peripheral nerve unit is formed, when being activated as neuron A and being flowed to neuron B mono-(or a string) stimulation by cynapse transmission, this critical stimulus makes the cynapse of neuron B integrate the threshold value finally reached needed for activation finally, and triggers generation action potential.So from phenomenon and result, we can see, when neuron A is activated, is transmitted by cynapse and integrate, making next neuron B be triggered and form activation, also namely have neuron A activation → neuron B to activate, this relation and result constitute specific corresponding relation between the two.
2.3 can obtain in neuroid with this, being expressed as of the memory relation that neuron A and neuron B is formed: " in a neural network, no matter whether neuron A and neuron B also sets up with other neurons synaptic connections, as long as there is so a kind of connection relationship in neural network, make when neuron A by other stimulate activate (input of stimulation) and produce action potential time, this neural network have and only have (uniquely) neuron B can activate by neuron A group and produce output, so, we just think, neuron A and neuron B sets up memory relation.Namely neuron A activation → neuron B activates ".
(note: here, neuron A and B may be single neuron, also may be one group of neuron, i.e. neuron pool, but on Macro-Functions, represent a certain information, convenient in order to describe, be still regarded as a neuron).
Applicant calls " uniqueness reflection " phenomenon of neuroid this, and it comprises the implication of two aspects: one is that the input stimulus of neuron A is enough to cause the activation of neuron B to export, otherwise nature cannot have reflection to export; Two be neuronic activation output caused by neuron A is unique one, and namely the activation of neuron B exports, if neuron A may cause multiple neuronic activation, then net result does not have neuron effectively to be activated, and causes effectively not reflecting output.
Same, the output that is activated of neuron B, also may become the input neuron of next process, jointly causes the activation of another neuron C to export with neuron A.That is: neuron A activation+neuron B activation → neuron C activates and exports.Can the like, down to the output that the neuronic chain type that appearance one is complicated activates.(this neuronic chain type activates, essence is the essence of the thinking of brain, namely correspond to a series of neurons of customizing messages in brain, produces a string activation chain owing to being sequentially activated, the chain type activity of this information corresponding to activation chain is the thinking of brain.This has relevant describing in another of applicant is about the application documents of thinking, because the relation with the application is little, does not describe here).
2.4 again in conjunction with 2.2 statement, if neuron A correspond to information A, neuron B correspond to information B, then neuron A and neuron B just achieves the memory to information A and information B.
So, explain the memory of brain to information from neuronic aspect, be: " neural network that brain is consisted of neuron and cynapse, to the multiple information (multiple neuron is by synergistic activation) inputted, be associated relation; And upper once re-enter some of them information (activate some of them neuron) time, a certain customizing messages is exported according to incidence relation reflection, (namely cause a certain neuron other in neuroid be activated output), this process is memory ".
This memory principle of brain is diverse with the recording principle of computer.Computer is information recorded successively to be independently kept on medium when inputting, and independently read successively during output, its information can be isolated record and read again.And the memory of brain must have two or more different information could form correlativity and incidence relation, single isolated uncorrelated information cannot form memory.When remember condition, when different input information (information A and information B) have activated different neurons (neuron A and neuron B), because this activation belongs at cell space or IAS nearly section outburst high threshold action potential, so the cynapse that can cause between them produces STDP plasticity, set up incidence relation (neuron A → neuron B) by synaptic plasticity or synaptic remodelling, this process is " note "; Afterwards, when a certain information (information A) be transfused to and make a certain or some neuron (neuron A) is activated again time, due to the incidence relation (neuron A → neuron B) of the Synaptic junction between them, another neuron (neuron B) is activated, and the activation of this neuron (neuron B), can reflect and export corresponding information (information B), this process is " recalling ".The neuronal activation of this process can be at the action potential of IAS section outburst far away Low threshold, does not cause synaptic plasticity, so again do not form memory effect.
From the Forming Mechanism of memory, " information A → information B " this information is wanted can reliably obtain (recalling) after formation memory again, key is the corresponding relation that neuron A and neuron B sets up, whether can make when neuron A activates again, have and only have neuron B by corresponding activation, also namely reappearing unique reflective relation of " neuron A activation → neuron B activates ".This both can realize by strengthening validity that cynapse transmitted between neuron A and neuron B, also can by weakening neuron A and B and transmit usefulness with other neuronic cynapses thus between the relative A of improve and B, cynapse transmitting validity to realize.So, usefulness enhancing (as LTP) is transmitted in cynapse can produce memory, cynapse is transmitted usefulness and is suppressed (as LTD) can produce memory too, between neuron, growth forms new effective Synaptic junction and can produce memory, and between neuron, original effective cynapse is weakened inefficacy and can produce memory too.The mankind are in the Childhood, and when forming memory fast and in a large number, cynapse quantity but significantly reduces, and estimate to be exactly the process forming this memory below.Conversely, (forgeing) is lost in the memory of " information A → information B ", that is to say that " neuron A activation → neuron B activates " this unique reflective relation is damaged, both may to be that validity that between neuron A and neuron B, cynapse is transmitted is destroyed caused for this, also may be that neuron A and B sets up new more effective cynapse transmission with other neurons thus the validity that between the relative A of reducing and B, cynapse is transmitted caused.And with this inference, the firmness of imformation memory, follow and also have relation for the neuronic quantity of corresponding prescribed information and projection relation: if need a lot of neuronic synergistic activations that another neuron just can be made to occur unique reflex activation, (also namely needing more feature just can limit a certain information), then this memory is relatively more firm, because the obviously more difficult interference being subject to other neuronic activities of this many-to-one reflective relation.
2.5 foregoings describe the essence of brain memory, the ultimate principle of neuron formation memory.So, at brain structural level, cerebral nerve network how to set up the input and output path with external information, and form the memory to information?
Brain is a complicated information handling system, and single just for the process of external information, this system just includes the stage such as perception, identification, record, process, output to information.Output is input to from information, brain presents multiple treatment mechanism: 1, the information Perception stage, in the perception stage to external information, type and the parameter of being carried out perception external information by sensory neurone, external information is decomposed into dissimilar parameter, (such as visual information just comprises the many kinds of parameters such as space, color, brightness, movement), be converted into the frequency of action potential and the change information of time of many group neurocyte again, also namely encoded various parameter by time encoding and frequency coding, 2, information cognitive phase, sensory neuron output information projects relevant intermediate treatment maincenter, (such as, the intermediate treatment maincenter of visual information is foreign journals and primary visual cortex), by the neuron space structure of high-sequential, by one group or several groups of neuronic synergistic activations that the convert information of frequency coding and time encoding is space diverse location, also i.e. space encoding, 3, the information record stage, the output of intermediate treatment maincenter projects the relevant range in senior sensory cortex, (such as, senior visual cortex), this projection is the projection of " many-one ", on intermediate treatment maincenter, multiple or many groups of neuronic activation just cause one or one group of neuronic activation on senior cortex, the integration of the multiple parameters to information is completed with this, multiple parameter is made finally to collectively define original external information again, and represent this external information by one or one group of specific neuronic activity, the external information through identifying is realized to be recorded as internal information.This internal information can be have recorded the image of certain object, sound or other features, if relate to word language, then also needs to project another further and corresponds to semantic senior cortex, complete the identification to semanteme and record.4, intermediate treatment stage, brain is obviously the most complicated most important function again to the intermediate treatment of internal information, this includes the memory of information and integrates reflection, wherein information integration and reflection be in fact also the process that brain carries out thinking and reaction, this integration and reflection are mainly integrated by cynapse on cortex and neuronic activation reflex action, realize in a serial fashion, (in addition, brain also comprises memory and the reflection process of action, mood, endocrine etc.).5, information exports the stage, and the output of brain comprises internal control and outsidely to export, outside export then mainly to comprise speak, write and various body action, these all need to be realized by muscular movement, and cerebellum is the main portions of Collaborative Control muscular movement.When brain (akrencephalon) exports, the neuronal activation of different spatial (also i.e. space encoding) information projection to cerebellum, in cerebellum, the neuron of similar high-sequential again receives the information of brain, and be converted to the explosion time of the action potential of other output neurons and the change of frequency by collaborative electric discharge, (time encoding and frequency coding), and the details such as direction, speed and strength of muscular movement is modulated with this, muscular movement is accurately completed.
From in broad terms, the processing stage that above-mentioned brain being to each of information, all there is the phenomenon of memory.Such as in perception and the identifying of information, neuron just forms the memory to the course of work by self-adaptation, habituation, compensation and FEEDBACK CONTROL etc., and carries out adaptability adjustment with this to perception and identifying.And in the output stage of information, the neuron of neuron especially in cerebellum is the coordination and the control procedure that are recorded various muscular movement by memory.But obviously at us in the research of imformation memory process, the most important thing is that the intrerneuron of brain is to the memory of internal information and reflection process, also i.e. the 2nd above-mentioned stage and information record stage, and the 3rd stage and intermediate treatment stage.
Wherein the 2nd stage and the information record stage, at this moment through the external information (as sound, image, sensation etc.) of Classification and Identification, its information coding is converted into space encoding from time encoding and frequency coding, that is to say and to encode the parameter of number of different types of external information by organizing neuronic activation more, then these neurons together project one or one group of specific neuron again, with this neuron maybe the neuronic activation of this group uniquely represent original external information.So in the memory of this one-phase to information, essence is that spatially different neuron is coordinated to activate owing to being constantly transfused to exciting signal, and sets up input and integrate the reflective relation exported.This memory generally can be realized by synaptic plasticity, and, due to the memory of this one-phase, much belong to multiple feature to limit an information, (such as, the image information of an object, just needs a lot of space characteristics, color characteristic, brightness etc. to limit), so its reflective relation formed is not easy to be damaged, that is to say that formed memory can be more firm.As for concrete memory forming process, can with reference to the first neurn simulation network (Figure 22) of the present invention, and the principle of work of the second neurn simulation network (Figure 23) and process.
And the intermediate treatment stage of the 3rd stage and information, being brain upgrades the further integration of the internal information after processing, restructuring, preservation and with this, new input information is carried out to the process of reactive reflection process (such as thinking).Which comprises at least the information processing of the several types such as declarative information, procedural acts, mood, by the length of the timeliness of its memory, be often divided into working memory, long-term memory (also having title short-term memory) and long-term memory (or claiming permanent memory).Wherein the timeliness of working memory is very short, in fact that neuron is carrying out information reflecting in the process of process (such as thinking), because cynapse integration and cynapse are modulated and spontaneous a kind of effect, if brain does not have subconsciousness to go to remember it, then working memory just disappears very soon.And the generation mechanism of long-term memory and long-term memory, then transmit plasticity (STDP plasticity) by cynapse and synaptic remodelling realizes, its concrete Forming Mechanism in neuron aspect can with reference to the principle of work of the second neurn simulation network of the present invention.Wherein how declarative memory is formed and conversion in the region such as cortex and hippocampus, carry out below describing.
Three, the difference remembered of working memory, long-term memory (hippocampus memory) and long-term (forever) and conversion.
Applicant summarizes in " cortex-hippocampus " information processing path, and the process remembered the memory of declarative input information realization, journey memory in short-term, long-term memory and permanent (for a long time) is as follows.(in fact hippocampus described here should be Medial Temporal Lobe, comprise the entorhinal cortex near hippocampus and hippocampus, smell all cortexes, other hippocampal cortex three cortical areas, more be familiar with due to the importance of hippocampus in memory, convenient in order to describe, just hippocampus is referred to as in these regions here).
3.1 when information inputs, such as input information " A is---", if be originally formed with the memory of " A is B " on cortex, namely cortex epineural unit " A " and neuron " B " have set up directly effective synaptic contact, then at this moment directly there is neuronic chain type to activate, " A " activates → " B " activation, produces recalling or thinking of information, directly reflects output information " B " when namely inputting information " A ".(will note neuron A and B here equally, may be single neuron, also may be one group of neuron and neuron pool, but on Macro-Functions, represent a certain information, convenient in order to describe, and is still regarded as a neuron).
If 3.2 inputs is declarative information " A is B ", and this information not yet exists on cortex, namely the related Neurons " A " on cortex and " B " do not set up effective synaptic contact, cannot produce the chain type activation that " A " activation → " B " activates, then the neuron from cortex is projected the neuron of hippocampus by bi-directional path and makes its excited activation by neuronic activation action.Neuron " A " on such as cortex activates the neuron " C " projected on hippocampus makes it activate, and " D " that " B " on cortex projects on hippocampus makes it activate; Because the neuron on hippocampus exists good synaptic plasticity (LTP and LDP), and activated by successively synchronous with " D " due to " C ", so set up incidence relation by STDP synaptic plasticity between the neuron of hippocampus " C " and " D ", namely " C " activates → " D " activation, produces memory relation.Certainly, this process may need repeatedly to stimulate just formation memory relation, and also once may stimulate and just be enough to realize, these needs are determined according to the relation between information.
3.3 like this, as input signal A again, namely when the information that inputs makes cortex epineural unit " A " be activated, this indirect pathway can be activated according to " A " activation → " C " activation → " D " activation → " B ", by the memory relation of the intrerneuron " C " on hippocampus and " D ", indirectly make " B " on cortex to be also activated, be also finally " A " activate → " B " activate, thus output information " B ", namely realize recalling of information " A being B ".This synaptic plasticity on hippocampus is powerful not when just starting to be formed, the memory formed is also not bery firm, so memory easily forgets loss in early days in formation, but synaptic plasticity constantly can be strengthened with repeatedly input stimulus, more firmly remember so information can produce through repeatedly input.Memory based ageing length, these memories can be divided into journey memory in short-term and long-term memory, journey memory in short-term and long-term memory are all be kept in hippocampus, rely on the synaptic plasticity of hippocampal neuron to realize, so this that is to say that so-called hippocampus is remembered.Hippocampus memory relies on synaptic plasticity to realize, due to synaptic plasticity to there is timeliness still shorter, is not again stimulated if long-time (recall or again input), still can be caused memory loss because synaptic plasticity weakens.
If 3.4 these processes are repeatedly repeated, namely repeatedly input or recall information " A is B ", make the neuron on cortex " A " and " B " because repeatedly excited being stimulated by synchronous, then neuron " A " and " B " can activate unborn " prefabricated property " cynapse between the two, " prefabricated property " cynapse is activated and is converted into " validity " cynapse possessing and effectively transmit usefulness, (this conversion process comprises the change of synaptic structure and component, need promotor gene to transcribe to synthesize with protein, even also may be grow the effective cynapse made new advances, in fact also be equivalent to a kind of synaptic remodelling), so set up directly effective synaptic contact between cortex epineural unit " A " and " B ", namely " A " activates → " B " direct path of activating." prefabricated property " cynapse between this cortex epineural unit is formed to the conversion of " validity " cynapse, should be depend on the synchronous stimulation that information repeatedly inputs or recall to realize, but seem also there is research explanation, brain is in some stage (such as rapid eye movement phase) of sleep, may occur that a lot of related Neurons is unconsciously activated, so also positive role may be produced to the direct path that formation " A " activation → " B " activates, that is, at sleep stage, this conversion may be promoted due to neuronic synchronous activation.
3.5 so far, when input signal " A " again, namely when the information that inputs makes cortex epineural unit " A " be activated, on cortex, then pass through the direct path that neuron " A " activation → " B " activates, neuron " B " is directly made to activate, recalling or thinking of generation information, directly reflects output information " B " when namely inputting information " A ", memory has transformed and has been kept on cortex.It is the difference depending on synaptic plasticity to realize with the memory on hippocampus, the memory of cortex depends on " prefabricated property " cynapse to activate and is converted into " validity " cynapse and realizes, the cynapse of " validity " is once be formed just highly stable, so the memory of cortex is formed slower, but it is once be formed, very firm, can long term maintenance effective, be called " permanent memory " or " long-term memory ".
3.6 when the memory path of cortex is once produce, activation transmission path due to them is shorter more direct, the information transmission of " A is B " then, just the indirect pathway on hippocampus is abandoned, carry out, so the maintenance of long-term memory no longer relies on hippocampus from neuron " A " direct path that activation → " B " activates cortex.And original on hippocampus " A " activation → " C " activation → " D " activation → " B " activates this and indirectly remembers path, can, reduction also ultimate failure suppressed owing to not reusing.Its inefficacy mechanism remains STDP effect, namely because the information transfering path on cortex is shorter more direct, information is from cortex epineural unit " A " to → successful transmission of " B ", will early than (faster than) from the indirect transfer of " A " hippocampus → " C " → " D " → " B ", the activation excitement which results in the terminal nerve unit of hippocampus indirect transfer path will early than the neuron of front end, so produce cynapse according to STDP mechanism to suppress, and the inefficacy of final whole piece path.Each neuron of the memory path of this inefficacy can discharge, for forming the indirect pathway of new hippocampus memory to the new input information of cortex, (each neuron is formed with synaptic contact with other a large amount of neurons), so the little hippocampus of volume can not be saturated by appearance because the continuous various fresh information of memory.
3.7 above-mentioned examples are that real work also may exist the participation of some bidirectional reflectance paths and modularity intrerneuron with the simplest information " A is B " and the neuromechanism that simplifies most so that the course of work to be described.And for the information of complexity, just according to the same reason, by the intermediate nerve metanetwork of the multi-level cross interconnected complexity in multi partition in hippocampus (comprising multiple region of interest such as other hippocampus), set up the indirect pathway of this hippocampus memory.Hippocampus relies on the good synaptic plasticity of its intrerneuron to set up mediate memory path, and what play in the forming process of declarative memory is a kind of similar intermediate axle beam action, and just as one " transfer shed ", cortex is then " long-term warehouse ".So, we finally, first various exactly declarative memory is stored on hippocampus, dumps on cortex afterwards unawares.This conversion process has relation with the length of time, and the frequent Degree of Accord Relation also with sign stimulus is close.And, before the direct path of cortex is set up, information transmission can use the indirect pathway of hippocampus always, after only having the direct path foundation of cortex, indirect pathway on hippocampus was just lost efficacy by reduction gradually, so the situation that unloading there will not be tomography is carried out in memory from hippocampus to cortex.
The formation of 3.8 these memories above-mentioned and long-term memory and shift to new management mechanisms can best explain: (1), hippocampus extracts and cannot form new long time-histories declarative memory.---because hippocampus is extractd cause the mediate memory path cannot setting up hippocampus, be therefore also just difficult to Stimulation of The Brain cortex and set up direct path.(2), hippocampus excision can cause the retrograde memory loss of time laminarity.---because memory transforms from hippocampus to cortex both needed the time, simultaneously also close with the frequent Degree of Accord Relation stimulated, the memory of different information transforms obviously can't be synchronous in time.When hippocampus is extractd, part memory has been transformed on cortex, and part memory is still kept on hippocampus, therefore causes the time laminarity remembered to drive in the wrong direction and loses memory.(3), long-term memory is kept on cortex, and hippocampus is extractd does not affect these memories.---because the memory direct path of cortex is once set up, memory does not just rely on the indirect pathway of hippocampus.This mechanism also perfectly can explain the symptom of multiple animal memory experiment and famous amnesiac patients H.M..
3.9 according to this mechanism, applicant infers, some declarative memory (between information easily occur incidence relation), if through the long-time stimulation continued repeatedly, also can without the need to this middle bridge of hippocampus, and directly in cortex, dependence repetitious stimulation related Neurons the most at last " prefabricated property " cynapse directly activates as " validity " cynapse, set up direct effective path to realize long-term memory.If revised further the T-shaped labyrinth of classics or the experiment of radial maze, this inference should be verified.In fact, in the symptom about famous amnesiac patients H.M. describes, although patient H.M. cannot remember the thing just occurred soon at one's side, can know clearly that himself is one and forgets patient, this example has also confirmed above-mentioned inference.It is said, when doing complicated memory problems to him, he can remind tester's " you know that I has memory disorders ".Although this illustrates that patient picked-off hippocampus, hippocampus memory cannot be formed, but " I has memory disorders " this information (belongs to typical declarative memory, and be just obtain after extracing hippocampus), still long-term memory can be formed in his cerebral cortex owing to repeatedly inputting, otherwise he should feel fresh to testing each time, also can not remember that oneself being one forgets patient.
3.10 applicant thinks, memory and long-term memory this formation and shift to new management mechanisms, have it rational for biosome in the information processing of vital movement process: first, it can form memory fast by hippocampus, can be processed by this memory as early as possible and tackle fresh information, but be not temporarily recorded in securely on cortex, (because also not confirming that whether these fresh informations are important and significant for more time), so also just can ensure the memory stability of some important informations, can not because of constantly inputting and constantly updating fresh information and cause excessive shock to original imformation memory structure and thoughtcast.And at fresh information by repeated multiple times input or after using, illustrate that these information are important, just naturally transform and be recorded in securely on cortex, incorporating original message structure, and long-term impact is produced on the thinking habit of biosome and behavioural habits.This also can explain some macroscopic appearances of human brain very well, and such as, to the study of new knowledge new things with accept very fast, and experience, idea, custom, personality, the mode of thinking, behavioural characteristic etc. are relatively stable, but can slowly change.
3.11 obviously, path is remembered to the formation of declarative memory with shift to new management mechanisms similar with " cortex-hippocampus ", " cortex-corpus straitum " is also had to remember path to the formation of procedural memory and conversion, and " cortex-amygdaloid nucleus " remember formation and conversion that path remembers emotionality, and these paths are formed each other and intermodulate.According to this mechanism, the zoopery of multiple procedural memory and emotional memory can be explained very well, also can simplicity of explanation experience and mood on the impact of thinking and behavior, or study thinking is on the impact of mood.Further, in brain, certain path should be also had, the endocrine material change information of biology is remembered and reflected.May to be the traditional Chinese medical science by some drugs adjust the constitutive character of human body thus reach the biology working mechanism for the treatment of or health-care effect for this.According to such supposition, in the traditional Chinese medical science, the constitutive character of human body (such as: empty, strong, hot, cold etc., the sign of blood pressure, blood fat, pathoglycemia and so on even), may be that the ratio of human body to some endocrine substance creates memory and habituation, and therefore make human organ form the duty adapted, if there is the change of related substances concentration due to external disturbance (environment, food or take medicine), then human body just may cause reflex response, there is symptom or mitigation symptoms, then also slowly can be suitable for (habituation) again.
3.12 applicants infer, although memory retention is more firm on cortex, but still may be weakened because of cynapse competition and even lose." prefabricated property " cynapse is activated as the process of " validity " cynapse, need more formation cynapses and realize cynapse transmit material (protein, mediator, nutriment etc.), this may absorb these materials of other cynapses of surrounding, thus have impact on the transmission usefulness of other cynapses original around, produce the phenomenon of cynapse competition.Such as: on effective path of neuron " A " → " B " set up on cortex, if newly established again " A " → " E ", " F " → " B " etc. are path effectively, because " validity " cynapse of these new activation is to the tight demand of the material formed needed for effective cynapse, may cause " A " → " B " path between the minimizing of necessary material of cynapse, thus making it transmit usefulness decline, cynapse weakens and even withers away.From macroscopically, be exactly after we have remembered information " A is B ", if remembered again " A is E ", " F is B " and and so on close to the information easily obscured, then not remembering of information " A is B " can be caused solid, even lose.And this, may also be that biosome realizes information updating thus retains a kind of instinct of the information of the most useful value.

Claims (10)

1. a neuron analogue means, includes dendron input end, signal processing module and aixs cylinder output terminal, signal processing module includes film integrating circuit, film discharge circuit, first threshold trigger circuit circuit, first action potential pulsing circuit, it is characterized in that: also include Second Threshold trigger circuit and the second action potential pulse-generating circuit, the input end of Second Threshold trigger circuit is connected to the output terminal of film integrating circuit, the output terminal of Second Threshold trigger circuit is connected to the second action potential pulse-generating circuit, the output terminal of the second action potential pulse-generating circuit connects aixs cylinder output terminal, the output terminal of the second action potential pulse-generating circuit is also connected to dendron input end by a reverse channel transmission simultaneously,
Between the input end of described first threshold trigger circuit and the output terminal of film integrating circuit, be added with a signal delay circuit; The output terminal of the second action potential pulse-generating circuit is also connected to film discharge circuit simultaneously;
Described neuron analogue means is also with " an attentional manipulation end " for switch operating state;
Described Second Threshold trigger circuit, also with threshold regulating circuit; The input control end of threshold regulating circuit, is connected to " attentional manipulation end ";
Between described " attentional manipulation end " and the input control end of threshold regulating circuit, also with the delay holding circuit of a duty.
2. a neuronic analogy method, comprising:
(1), integration is carried out to the signal of dendron input end input;
(2), detect the voltage signal after integration;
(3) if this voltage signal is less than setting value 1, then neuron does not produce action; If this voltage signal is equal to or greater than setting value 1 but its voltage rise slope is less than setting value 2, then trigger and produce action potential pulse, output to aixs cylinder output terminal, the voltage signal after integration is reset simultaneously; If this voltage signal is equal to or greater than setting value 1 and its voltage rise slope is equal to or greater than setting value 2, then trigger and produce action potential pulse, not only output to aixs cylinder output terminal but also oppositely output to dendron input end simultaneously, the voltage signal after integration is reset simultaneously;
Described neuron analogy method also switches memory and reflection two kinds of different operating states by the signal condition of " attentional manipulation end ": when " attentional manipulation end " signal condition is for " reflection ", neuron carries out work by the method that step is (1) extremely (3) described; When " attentional manipulation end " signal condition is " memory ", (3) step the changes into: if this voltage signal is less than setting value 1, then neuron does not produce action; If this voltage signal is equal to or greater than setting value 1, then neuron triggers and produces action potential pulse, not only outputs to aixs cylinder output terminal but also oppositely output to dendron input end simultaneously, is reset by the voltage signal after integration simultaneously.
3. the second neuron analogue means, includes dendron input end, signal processing module, aixs cylinder output terminal; Signal processing module includes film integrating circuit, film discharge circuit, first threshold trigger circuit and the first action potential pulse-generating circuit, it is characterized in that: described neuron analogue means is also with action potential reverse channel transmission; The output terminal of the first action potential pulse-generating circuit, is connected to dendron input end by this reverse channel transmission; The break-make of reverse channel transmission is subject to the control of a reverse transfer control circuit; The control input end of reverse transfer control circuit, is connected to one for switching " the attentional manipulation end " of neuron memory/reflective operation state;
Between described " attentional manipulation end " and the control input end of reverse transfer control circuit, be also provided with the delay holding circuit of a duty;
Described neuron analogue means also includes Second Threshold trigger circuit, and the input end of Second Threshold trigger circuit is connected to the output terminal of film integrating circuit, and the output terminal of Second Threshold trigger circuit is connected to the control input end of reverse transfer control circuit;
Between the input end of described first threshold trigger circuit and the output terminal of film integrating circuit, be provided with a signal delay circuit;
Between the output terminal of described Second Threshold trigger circuit and the control input end of reverse transfer control circuit, be also provided with a Time delay holding circuit.
4. the second neuron analogy method, comprising:
(1), integration is carried out to the signal of dendron input end input;
(2), detect the voltage signal after integration, if this voltage signal is less than setting value 1, then do not produce action;
If the voltage signal (3) after integration is equal to or greater than setting value 1, then detect the signal condition of " attentional manipulation end ";
If " attentional manipulation end " signal condition is " reflection " state, then neuron fires pulse, outputs to aixs cylinder output terminal, the voltage signal values after integration is reset simultaneously;
If " attentional manipulation end " signal condition is " memory " state, then neuron fires pulse, had both outputted to aixs cylinder output terminal, oppositely outputted to dendron input end again, the voltage signal values after integration was reset simultaneously;
Described " attentional manipulation end " is also with the delay holding circuit of a duty;
Described the second analogy method also comprises further:
(4), detect the integral voltage signal after integration, if this voltage signal its voltage rise slope while being greater than setting value 1 is also greater than setting value 2, then neuron produces an action potential pulse, has both outputted to aixs cylinder output terminal, oppositely outputs to dendron input end again; Voltage signal values after integration is reset simultaneously.
5. " solidification cynapse " analogue means, comprises cephacoria input end, caudacoria output terminal, it is characterized in that: be serially connected with a capacitor discharging circuit between cephacoria input end and caudacoria output terminal, and capacitor discharging circuit simultaneously and be connected to an electric capacity leadage circuit;
Described capacitor discharging circuit is made up of a discharge capacity, and the front end of discharge capacity connects cephacoria input end, and the rear end of discharge capacity connects caudacoria output terminal by one for setting the transmission resistance transmitting usefulness; Transmit the two ends of resistance, reverse parallel connection has a charging diode.
6. the analogue means of " modifiable synapse ", comprises cephacoria input end, caudacoria output terminal, STDP chronicle recognition circuit and transmits usefulness Circuit tuning; It is characterized in that: STDP chronicle recognition circuit is with LTP output terminal and LTD output terminal; LTP output terminal and LTD output terminal are connected to an accumulate device for store electrical energy respectively by a charge circuit and a discharge loop; Accumulate device is connected with an initialization charging circuit; The output of accumulate device is connected to cynapse and transmits usefulness Circuit tuning, and the synchro control that usefulness Circuit tuning accepts the input signal of cephacoria input end is simultaneously transmitted in cynapse;
Described accumulate device is also connected with a two-way leadage circuit; Described two-way leadage circuit comprises the mu balanced circuit that has setting voltage, and mu balanced circuit is connected on accumulate device by a resistance;
Described STDP chronicle recognition circuit all has threshold test to the signal that cephacoria input end and caudacoria output terminal input, and only the action potential spike higher to voltage amplitude identifies.
7. the analogue means of " cynapse of prefabricated property ", the cynapse transmission channels including cephacoria input end, caudacoria output terminal and preset; It is characterized in that: also with a STDP chronicle recognition circuit, for identifying the spike signal between cephacoria input end and caudacoria input end with STDP feature; The output of STDP chronicle recognition circuit is connected to a STDP integrating circuit, and the output of STDP integrating circuit is connected to a thresholding control circuit; The output terminal of thresholding control circuit is connected to a channel switching circuit; Channel switching circuit is for controlling the break-make of the cynapse transmission channels preset;
Described default cynapse transmission channels is a cynapse transmission channels having that usefulness is transmitted in cynapse;
The course of work of described " cynapse of prefabricated property " analogue means is: STDP integrating circuit according to setting rule, accumulate STDP chronicle recognition circuit within a certain period of time identify the signal of output, and be converted into the output valve of output signal; Thresholding control circuit, for detecting the output valve of the output signal of integrating circuit, reaches when the output valve of integrating circuit or higher than setting thresholding 1, then produces and export control signal; And start connection cynapse transmission channels by channel switching circuit, set up between cephacoria input end and caudacoria output terminal and there is the cynapse transmission channels that usefulness is transmitted in cynapse;
Its course of work further comprises: after channel switching circuit starts connection cynapse transmission channels, thresholding control circuit continues the output valve of the output signal detecting integrating circuit, when the output valve of integrating circuit is lower than setting thresholding 2, then closes and export control signal; And close cynapse transmission channels by channel switching circuit;
The setting rule of described integrating circuit is: the output valve improving output signal when there is the signal of LTP feature, reduces the output valve of output signal when there is the signal of LTD feature.
8. one according to claim 7 " cynapse of prefabricated property " analogue means, is characterized in that: described STDP integrating circuit includes charging circuit, discharge circuit and accumulate device; The LTP output terminal of STDP chronicle recognition circuit and LTD output terminal are connected on accumulate device respectively by charging circuit and discharge circuit; The output terminal of accumulate device is connected to the input end of thresholding control circuit;
The STDP characteristic control LTP output terminal of spike signal that described STDP chronicle recognition circuit occurs according to cephacoria input end and caudacoria output terminal and the output signal of LTD output terminal; STDP chronicle recognition circuit LTP output terminal when identifying LTP feature occurs that pulse exports, and is charged to accumulate device by charge circuit; STDP chronicle recognition circuit LTD output terminal when identifying LTD feature occurs that pulse exports, and is discharged to accumulate device by discharge loop;
Thresholding control circuit is used for detecting the output voltage of accumulate device, when the output voltage of accumulate device meets or exceeds setting threshold value 1, thresholding control circuit exports control signal, and start connection cynapse transmission channels by channel switching circuit, set up between cephacoria input end and caudacoria output terminal and there is the cynapse transmission channels that usefulness is transmitted in cynapse;
After channel switching circuit starts connection cynapse transmission channels, thresholding control circuit continues the output voltage detecting accumulate device, when the output voltage of accumulate device is lower than setting thresholding 2, then closes output control signal; And close cynapse transmission channels by channel switching circuit.
9. the neurn simulation network adopting neuron analogue means of the present invention to form, comprise at least one neuron display unit, each neuron display unit comprises multiple presynaptic neuron analogue means and multiple postsynaptic neuron analogue means, between the aixs cylinder output terminal of presynaptic neuron analogue means and the dendron input end of postsynaptic neuron analogue means, employing has the plastic cynapse analogue means of cynapse transmission usefulness STDP and connects, it is characterized in that: " the attentional manipulation end " of each neuron analogue means links together, controlled by the control signal of a road for the memory/reflection two kinds of duties switching neuron analogue means.
10. there is a neurn simulation network for two kinds of different time course memories, comprise the first input neuron group, the first output neuron group, the second input neuron group, the second output neuron group, the first intrerneuron group and the second intrerneuron group; It is characterized in that:
The aixs cylinder output terminal of the first input neuron group is connected with adopting " solidification cynapse " between the dendron input end of the first output neuron group, and the aixs cylinder output terminal of the second input neuron group is connected with adopting " solidification cynapse " between the dendron input end of the second output neuron group;
The aixs cylinder output terminal of the first input neuron group and the second input neuron group, and between the dendron input end of the first intrerneuron group, adopt " modifiable synapse " to connect;
The aixs cylinder output terminal of the first intrerneuron group, and between the dendron input end of the second intrerneuron group, adopt " modifiable synapse " to connect;
The aixs cylinder output terminal of the second intrerneuron group, and between the dendron input end of the first output neuron group and the second output neuron group, adopt " modifiable synapse " to connect;
Between the aixs cylinder output terminal of the first input neuron group and the dendron input end of the second output neuron group, between the aixs cylinder output terminal also having the second input neuron group and the dendron input end of the first output neuron group, " cynapse of prefabricated property " is adopted to connect.
CN201410606697.7A 2014-07-21 2014-10-30 The simulator and method of neural network Active CN104899640B (en)

Priority Applications (5)

Application Number Priority Date Filing Date Title
CN201910669674.3A CN111291873A (en) 2014-07-21 2014-10-30 Method and device for simulating prefabricated synapse
CN201410606697.7A CN104899640B (en) 2014-07-21 2014-10-30 The simulator and method of neural network
CN201910669575.5A CN111275178A (en) 2014-07-21 2014-10-30 Neuron simulation method and device with different working forms
CN201710012154.6A CN107045649B (en) 2014-07-21 2014-10-30 Simulated neural network with short-term and long-term memory
CN201910694308.3A CN111353590A (en) 2014-07-21 2014-10-30 Synapse simulation method and device with STDP synapse plasticity

Applications Claiming Priority (3)

Application Number Priority Date Filing Date Title
CN201410347070 2014-07-21
CN2014103470704 2014-07-21
CN201410606697.7A CN104899640B (en) 2014-07-21 2014-10-30 The simulator and method of neural network

Related Child Applications (4)

Application Number Title Priority Date Filing Date
CN201710012154.6A Division CN107045649B (en) 2014-07-21 2014-10-30 Simulated neural network with short-term and long-term memory
CN201910669674.3A Division CN111291873A (en) 2014-07-21 2014-10-30 Method and device for simulating prefabricated synapse
CN201910669575.5A Division CN111275178A (en) 2014-07-21 2014-10-30 Neuron simulation method and device with different working forms
CN201910694308.3A Division CN111353590A (en) 2014-07-21 2014-10-30 Synapse simulation method and device with STDP synapse plasticity

Publications (2)

Publication Number Publication Date
CN104899640A true CN104899640A (en) 2015-09-09
CN104899640B CN104899640B (en) 2019-09-10

Family

ID=54032293

Family Applications (5)

Application Number Title Priority Date Filing Date
CN201910669575.5A Pending CN111275178A (en) 2014-07-21 2014-10-30 Neuron simulation method and device with different working forms
CN201710012154.6A Active CN107045649B (en) 2014-07-21 2014-10-30 Simulated neural network with short-term and long-term memory
CN201910669674.3A Pending CN111291873A (en) 2014-07-21 2014-10-30 Method and device for simulating prefabricated synapse
CN201410606697.7A Active CN104899640B (en) 2014-07-21 2014-10-30 The simulator and method of neural network
CN201910694308.3A Pending CN111353590A (en) 2014-07-21 2014-10-30 Synapse simulation method and device with STDP synapse plasticity

Family Applications Before (3)

Application Number Title Priority Date Filing Date
CN201910669575.5A Pending CN111275178A (en) 2014-07-21 2014-10-30 Neuron simulation method and device with different working forms
CN201710012154.6A Active CN107045649B (en) 2014-07-21 2014-10-30 Simulated neural network with short-term and long-term memory
CN201910669674.3A Pending CN111291873A (en) 2014-07-21 2014-10-30 Method and device for simulating prefabricated synapse

Family Applications After (1)

Application Number Title Priority Date Filing Date
CN201910694308.3A Pending CN111353590A (en) 2014-07-21 2014-10-30 Synapse simulation method and device with STDP synapse plasticity

Country Status (1)

Country Link
CN (5) CN111275178A (en)

Cited By (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105279557A (en) * 2015-11-13 2016-01-27 徐志强 Memory and thinking simulation device based on human brain working mechanism
CN105426957A (en) * 2015-11-06 2016-03-23 兰州理工大学 Neuron electrical activity simulator under electromagnetic radiation
CN107480597A (en) * 2017-07-18 2017-12-15 南京信息工程大学 A kind of Obstacle Avoidance based on neural network model
CN107688853A (en) * 2016-08-05 2018-02-13 北京中科寒武纪科技有限公司 A kind of device and method for being used to perform neural network computing
WO2018094717A1 (en) * 2016-11-28 2018-05-31 华为技术有限公司 Signal processing method and device based on spiking neural network
CN108875919A (en) * 2017-07-21 2018-11-23 北京旷视科技有限公司 Data processing equipment, data processing method and storage medium product
CN109102071A (en) * 2018-08-07 2018-12-28 中国科学院微电子研究所 A kind of neuron circuit and nerve network circuit
CN109255430A (en) * 2018-07-12 2019-01-22 电子科技大学 A kind of neuron coding circuit
CN109374021A (en) * 2018-10-30 2019-02-22 广州发展集团股份有限公司 The anti-interference acquisition system of physical quantity, method, apparatus and computer equipment
CN109460785A (en) * 2018-10-24 2019-03-12 广东工业大学 A method of the spike real-time grading accelerated based on GPU
CN110097171A (en) * 2019-06-19 2019-08-06 吉林大学 A kind of activity trajectory localization method and system based on scorpion micro-vibration Location Mechanism
CN110751273A (en) * 2018-07-22 2020-02-04 徐志强 Neuron and synapse simulation assembly
CN110931123A (en) * 2019-12-06 2020-03-27 深圳市德力凯医疗设备股份有限公司 Matrix brain network and construction method thereof
CN110991628A (en) * 2019-11-02 2020-04-10 复旦大学 Neuron circuit based on charge pump
CN111433790A (en) * 2017-11-15 2020-07-17 汉阳大学校产学协力团 Pulse driving apparatus for minimizing asymmetry of weight of synapse device and method thereof
CN112101535A (en) * 2020-08-21 2020-12-18 中国科学院深圳先进技术研究院 Signal processing method of pulse neuron and related device
CN112149813A (en) * 2020-09-18 2020-12-29 明峰医疗系统股份有限公司 Detector integral capacitance dynamic regulation method based on neural network
CN113767402A (en) * 2019-04-29 2021-12-07 ams国际有限公司 Computationally efficient implementation of simulated neurons

Families Citing this family (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109191498B (en) * 2018-09-05 2021-04-02 中国科学院自动化研究所 Target detection method and system based on dynamic memory and motion perception
CN109325596B (en) * 2018-09-17 2024-03-26 中国传媒大学 Nerve synapse plasticity calculation method based on calcium concentration
CN109886384B (en) * 2019-02-15 2021-01-05 北京工业大学 Bionic navigation method based on mouse brain hippocampus grid cell reconstruction
CN110263924B (en) * 2019-06-19 2021-08-17 北京计算机技术及应用研究所 Parameter and state estimation method of neuron population model
CN110390386B (en) * 2019-06-28 2022-07-29 南京信息工程大学 Sensitive long-short term memory method based on input change differential
CN110472726B (en) * 2019-07-25 2022-08-02 南京信息工程大学 Sensitive long-short term memory method based on output change differential
CN110490299B (en) * 2019-07-25 2022-07-29 南京信息工程大学 Sensitive long-short term memory method based on state change differential
KR102459296B1 (en) * 2020-07-22 2022-10-27 울산과학기술원 Method and apparatus for extracting neuronal networking properties related to kinematic parameters for decoding process
CN112016684B (en) * 2020-08-06 2022-09-02 广西大学 Electric power terminal fingerprint identification method of deep parallel flexible transmission network
CN112232440B (en) * 2020-11-10 2022-11-11 北京理工大学 Method for realizing information memory and distinction of impulse neural network by using specific neuron groups
CN112819142B (en) * 2021-02-04 2024-01-19 成都市深思创芯科技有限公司 Short-time synaptic plasticity work memory computing system and method
CN113658493B (en) * 2021-08-20 2023-05-02 安徽大学 Reinforced learning bionic circuit architecture for simulating associative memory

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101770560A (en) * 2008-12-31 2010-07-07 南方医科大学 Information processing method and device for simulating biological neuron information processing mechanism
CN101771412A (en) * 2008-12-31 2010-07-07 南方医科大学 Analog-digital conversion method and device based on neuron working principle
CN103733209A (en) * 2011-08-16 2014-04-16 高通股份有限公司 Method and apparatus for neural temporal coding, learning and recognition

Family Cites Families (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CA893371A (en) * 1972-02-15 W. Rork Danny Associative memory systems
FI894021A (en) * 1988-08-31 1990-03-01 Fujitsu Ltd NEURONSTRUKTUR.
WO1991018349A1 (en) * 1990-05-22 1991-11-28 International Business Machines Corporation Scalable flow virtual learning neurocomputer
JP3095798B2 (en) * 1991-03-13 2000-10-10 科学技術振興事業団 Analog back-propagation learning circuit and its driving method
US5438646A (en) * 1992-08-19 1995-08-01 Nec Electronics, Inc. Feed-forward neural network
US5864835A (en) * 1997-03-24 1999-01-26 Gorelik; Vladimir A. Apparatus for simulating a biological neuron
EP1182609A1 (en) * 2000-08-25 2002-02-27 Peter Bezler Circuit for the simulation of neural networks and for position-independent pattern reognition
US6754645B2 (en) * 2001-03-20 2004-06-22 Winbond Electronics Corp. Voltage-mode pulse width modulation VLSI implementation of neural networks
US7818273B2 (en) * 2007-09-18 2010-10-19 International Business Machines Corporation System and method for cortical simulation
US8250011B2 (en) * 2008-09-21 2012-08-21 Van Der Made Peter A J Autonomous learning dynamic artificial neural computing device and brain inspired system
US8447714B2 (en) * 2009-05-21 2013-05-21 International Business Machines Corporation System for electronic learning synapse with spike-timing dependent plasticity using phase change memory
US8515885B2 (en) * 2010-10-29 2013-08-20 International Business Machines Corporation Neuromorphic and synaptronic spiking neural network with synaptic weights learned using simulation
CN102496385B (en) * 2011-12-26 2014-04-16 电子科技大学 Spike timing activity conversion circuit
CN102610274B (en) * 2012-04-06 2014-10-15 电子科技大学 Weight adjustment circuit for variable-resistance synapses
CN103078055B (en) * 2013-01-04 2015-06-03 华中科技大学 Unit, device and method for simulating biological neuronal synapsis
CN103078054B (en) * 2013-01-04 2015-06-03 华中科技大学 Unit, device and method for simulating biological neuron and neuronal synapsis
CN103530690A (en) * 2013-10-31 2014-01-22 中国科学院上海微系统与信息技术研究所 Nerve cell element and neural network

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101770560A (en) * 2008-12-31 2010-07-07 南方医科大学 Information processing method and device for simulating biological neuron information processing mechanism
CN101771412A (en) * 2008-12-31 2010-07-07 南方医科大学 Analog-digital conversion method and device based on neuron working principle
CN103733209A (en) * 2011-08-16 2014-04-16 高通股份有限公司 Method and apparatus for neural temporal coding, learning and recognition

Cited By (32)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105426957A (en) * 2015-11-06 2016-03-23 兰州理工大学 Neuron electrical activity simulator under electromagnetic radiation
CN105426957B (en) * 2015-11-06 2017-09-29 兰州理工大学 A kind of electrical activity of neurons simulator under electromagnetic radiation
CN105279557A (en) * 2015-11-13 2016-01-27 徐志强 Memory and thinking simulation device based on human brain working mechanism
CN105279557B (en) * 2015-11-13 2022-01-14 徐志强 Memory and thinking simulator based on human brain working mechanism
CN107688853A (en) * 2016-08-05 2018-02-13 北京中科寒武纪科技有限公司 A kind of device and method for being used to perform neural network computing
WO2018094717A1 (en) * 2016-11-28 2018-05-31 华为技术有限公司 Signal processing method and device based on spiking neural network
US11308386B2 (en) 2016-11-28 2022-04-19 Huawei Technologies Co., Ltd. Signal processing method and apparatus based on spiking neural network
CN107480597A (en) * 2017-07-18 2017-12-15 南京信息工程大学 A kind of Obstacle Avoidance based on neural network model
CN107480597B (en) * 2017-07-18 2020-02-07 南京信息工程大学 Robot obstacle avoidance method based on neural network model
CN108875919A (en) * 2017-07-21 2018-11-23 北京旷视科技有限公司 Data processing equipment, data processing method and storage medium product
CN108875919B (en) * 2017-07-21 2021-04-02 北京旷视科技有限公司 Data processing apparatus, data processing method, and storage medium product
CN111433790B (en) * 2017-11-15 2023-09-19 三星电子株式会社 Pulse driving apparatus and method for minimizing asymmetry of weights of synaptic devices
CN111433790A (en) * 2017-11-15 2020-07-17 汉阳大学校产学协力团 Pulse driving apparatus for minimizing asymmetry of weight of synapse device and method thereof
CN109255430A (en) * 2018-07-12 2019-01-22 电子科技大学 A kind of neuron coding circuit
CN109255430B (en) * 2018-07-12 2022-03-15 电子科技大学 Neuron coding circuit
CN110751273A (en) * 2018-07-22 2020-02-04 徐志强 Neuron and synapse simulation assembly
CN109102071A (en) * 2018-08-07 2018-12-28 中国科学院微电子研究所 A kind of neuron circuit and nerve network circuit
CN109460785B (en) * 2018-10-24 2021-09-24 广东工业大学 Spike potential real-time classification method based on GPU acceleration
CN109460785A (en) * 2018-10-24 2019-03-12 广东工业大学 A method of the spike real-time grading accelerated based on GPU
CN109374021B (en) * 2018-10-30 2020-10-30 广州发展集团股份有限公司 Anti-interference physical quantity acquisition system, method and device and computer equipment
CN109374021A (en) * 2018-10-30 2019-02-22 广州发展集团股份有限公司 The anti-interference acquisition system of physical quantity, method, apparatus and computer equipment
CN113767402A (en) * 2019-04-29 2021-12-07 ams国际有限公司 Computationally efficient implementation of simulated neurons
CN110097171A (en) * 2019-06-19 2019-08-06 吉林大学 A kind of activity trajectory localization method and system based on scorpion micro-vibration Location Mechanism
CN110097171B (en) * 2019-06-19 2020-04-21 吉林大学 Activity track positioning method and system based on scorpion micro-vibration positioning mechanism
CN110991628B (en) * 2019-11-02 2023-04-18 复旦大学 Neuron circuit based on charge pump
CN110991628A (en) * 2019-11-02 2020-04-10 复旦大学 Neuron circuit based on charge pump
CN110931123A (en) * 2019-12-06 2020-03-27 深圳市德力凯医疗设备股份有限公司 Matrix brain network and construction method thereof
CN110931123B (en) * 2019-12-06 2024-01-05 深圳市德力凯医疗设备股份有限公司 Matrix type brain network and construction method thereof
CN112101535A (en) * 2020-08-21 2020-12-18 中国科学院深圳先进技术研究院 Signal processing method of pulse neuron and related device
CN112101535B (en) * 2020-08-21 2024-04-09 深圳微灵医疗科技有限公司 Signal processing method of impulse neuron and related device
CN112149813A (en) * 2020-09-18 2020-12-29 明峰医疗系统股份有限公司 Detector integral capacitance dynamic regulation method based on neural network
CN112149813B (en) * 2020-09-18 2022-05-13 明峰医疗系统股份有限公司 Detector integral capacitance dynamic adjustment method based on neural network

Also Published As

Publication number Publication date
CN107045649A (en) 2017-08-15
CN107045649B (en) 2020-06-02
CN104899640B (en) 2019-09-10
CN111275178A (en) 2020-06-12
CN111353590A (en) 2020-06-30
CN111291873A (en) 2020-06-16

Similar Documents

Publication Publication Date Title
CN104899640A (en) Simulator and method for neural network
Magee et al. Synaptic plasticity forms and functions
Gagliano et al. Experience teaches plants to learn faster and forget slower in environments where it matters
Sharma Interdependence: biology and beyond
Gordon Integrative neuroscience: Bringing together biological, psychological and clinical models of the human brain
Feinberg et al. The evolutionary and genetic origins of consciousness in the Cambrian Period over 500 million years ago
John et al. Anatomy and computational modeling of networks underlying cognitive-emotional interaction
Dayan et al. Uncertainty and learning
Olalde Rangel The systemic theory of living systems and relevance to CAM
CN104888349A (en) Method and device for carrying out brain activation on dead or brain-dead cerebrum
Esposito The obsolete self: Philosophical dimensions of aging
Costa et al. Dopamine
CN105279557A (en) Memory and thinking simulation device based on human brain working mechanism
Downey et al. Evolution and the brain
Gerstner Hebbian learning and plasticity
Naskar What is Mind?
Andrew The differential roles of right and left sides of the brain in memory formation
Yanagihara et al. Social interaction with a tutor modulates responsiveness of specific auditory neurons in juvenile zebra finches
Yuste et al. Neuronal ensembles: Building blocks of neural circuits
Li Computational models of the amygdala in acquisition and extinction of conditioned fear
Malaia et al. Salience of low-frequency entrainment to visual signal for classification points to predictive processing in sign language. In proceedings of 30th annual computational neuroscience meeting: CNS* 2021
Minne Interactions between synaptic plasticity and switches in brain states for memory consolidation: a modeling study
Rubin et al. A dynamic architecture of life
Wehle et al. Research on human cognition for biologically inspired developments: Human-robot interaction by biomimetic AI
Stöber Cooperate to compete: Identifying a potential role for hippocampal region CA2 in episodic memory formation

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant