CN104823205B - For enhancing the neural model of study - Google Patents
For enhancing the neural model of study Download PDFInfo
- Publication number
- CN104823205B CN104823205B CN201380063033.4A CN201380063033A CN104823205B CN 104823205 B CN104823205 B CN 104823205B CN 201380063033 A CN201380063033 A CN 201380063033A CN 104823205 B CN104823205 B CN 104823205B
- Authority
- CN
- China
- Prior art keywords
- channel
- tuple
- neuron
- nerve
- cynapse
- 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.)
- Expired - Fee Related
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
Abstract
It is a kind of for enhance study and selection of taking action neural model, the neural model includes multiple channels, the input in each channel neural tuple, the output nerve tuple in each channel, every group of input neuron in each channel is couple to every group of output neuron in each channel, further includes the award nerve tuple in each channel.The output neuron that the neural tuple of the award in each channel receives the input inputted from environment, and is coupled only in channel belonging to award neuron.If channel environment input be it is positive, the output neuron of respective channel award and have enhancing response, otherwise, the output neuron of respective channel it is penalized and have decay response.
Description
Cross reference to related applications
This application involves submit on December 3rd, 2012 application No. is 61/732,590 U.S. Provisional Patent Applications, and
It is required that it is as priority, this paper hereby incorporated by reference in its entirety.The application further relates to what on May 16th, 2013 submitted
Application No. is 13/896,110 U.S. Non-provisional Patent applications, and require it as priority, by reference that it is complete herein
Portion is incorporated herein.
The statement subsidized about federal government
The present invention is in government contract " the cynapse HR0011-09-C- of US Department of Defense Advanced Research Projects Agency neuron
It is carried out under 0001 " support.Government has certain right in the present invention.
Technical field
This disclosure relates to which neural network, is especially able to carry out the neural network of action selection and enhancing study.It is public herein
The technology opened includes the plastic action selection network for neuromorphic hardware.
Background technique
In the prior art, the neural network for being able to carry out selection of taking action has shown fine feature, has
Feature illustrates enhancing study.But in the prior art, action selection and enhancing learning algorithm are in for tip award problem
Reveal complicated solution, this is not easily controlled hardware embodiments.
The international neural network joint conference of Barr, D., P.Dudek, J.Chambers and K.Gurney in August, 2007
(IJCNN) " the Implementation of multi-layer leaky integrator of 1560-1565 pages
The basal ganglia on a neuron processor array is described in networks on a cellular processor array "
The model of section.Software neural model is able to carry out action selection.But Barr et al. is without describing about any of enhancing study
Inherent mechanism, and the microchannel of basal ganglion is predefined.
Merol la, P., J.Arthur, F.Akopyan, N.Imam, R.Manohar and D.Modha are in Electrical and Electronic
1-4 pages of " A digital neurosynaptic of the customization integrated circuit meeting (CICC) of Association of Engineers in September, 2011
One is described in core using embedded crossbar memory with 45pj per spike in 45nm "
The neuromorphic processor of table tennis game can be carried out with human opponent.But the network constructs offline, and once
Programming can keep static on hardware.
It is desirable that one kind can be realized action selection and enhancing learns and can be easier to hard-wired nerve net
Network.Multiple embodiments of the disclosure give answer to these and other demand.
Summary of the invention
It is a kind of for enhancing study and the neural model of selection of taking action includes: in one embodiment disclosed herein
Multiple channels；Input nerve tuple in each channel；Output nerve tuple in each channel, in each channel
Every group of input neuron be couple to every group of output neuron in each channel；Award neuron in each channel
Group, wherein every group of award neuron receives the input inputted from environment, and the wherein award neuron coupling in each channel
The output neuron being connected in channel belonging to award neuron；Wherein, if the environment input in channel be it is positive, it is corresponding logical
The output neuron in road is awarded and has the response of enhancing；Wherein, if the environment input in channel is negative, respective channel
Output neuron it is penalized and have decaying response.
It is a kind of for enhancing study and the neural model of selection of taking action includes: in another embodiment disclosed herein
Multiple channels；Input nerve tuple in each channel；Output nerve tuple in each channel, in each channel
Every group of input neuron be couple to every group of output neuron in each channel；Award neuron in each channel
Group, wherein every group of award neuron receives the input inputted from environment, and the wherein award neuron coupling in each channel
The output neuron being connected in channel belonging to award neuron；Inhibition nerve tuple in each channel, wherein often
Group inhibits neuron to receive the input for inhibiting the output nerve tuple in same channel belonging to neuron from the group, and its
In the neural tuple of inhibition in a channel have it is other each of other than the affiliated channel of inhibition nerve tuple to being located at
The output of output neuron in channel；Wherein, if the environment input of tuple neural for the award in channel be it is positive,
Then the output nerve tuple of respective channel is awarded and has the response of enhancing；Wherein, if the award for a channel is refreshing
Through tuple environment input be it is negative, then the output nerve tuple of respective channel it is penalized and have decaying response.
In another embodiment disclosed herein, a kind of basal ganglion neural network model includes: multiple channels；Position
Cortical neurons tuple in each channel；Striatal neuron group in each channel, every group of line in each channel
Shape somatic nerves member is couple to every group of cortical neuron in each channel；Award nerve tuple in each channel, wherein
Every group of award neuron receives the input inputted from environment, and wherein the award neuron in each channel is coupled only to award
Striatal neuron in channel belonging to neuron；The neural tuple of Substantia nigra reticulata (SNr) in each channel, wherein
Every group of SNr neuron is coupled only to the striatal neuron group in channel belonging to SNr neuron；Wherein, if for one
The environment input of the neural tuple of award in channel be it is positive, then the striatal neuron group of respective channel award and is had and enhanced
Response；Wherein, if the environment input of tuple neural for the award in a channel is negative, the corpus straitum of respective channel
Neural tuple is penalized and has the response of decaying；Wherein every group of SNr neuron is tatanic movable and by SNr neuron
The inhibition of striatal neuron in affiliated channel is incoming to be inhibited.
These and other feature and advantage can be clearly shown with attached drawing by the following detailed description.In attached drawing and say
In bright book, label indicates different features, refers to similar feature in the whole instruction label similar in attached drawing.
Detailed description of the invention
Fig. 1 shows a neural network according to the disclosure；
Fig. 2 shows another neural networks with lateral inhibition according to the disclosure；
Fig. 3 shows the basal ganglion neural network according to the disclosure；
Fig. 4 A to 4C shows the example according to one of disclosure enhancing study script；
Fig. 5 A to 5F shows the example of the synapse weight of the neural network according to the disclosure；
Fig. 6 is the schematic diagram for showing a kind of virtual environment of table tennis ball-type according to the disclosure；
Fig. 7-9 illustrate according to the virtual environment of the table tennis ball-type of Fig. 6 of the disclosure for different space width and when
Between span result；
Figure 10 illustrates the overall accuracy for the model for being 0.025 according to the space width of the disclosure.
Specific embodiment
In the following description, it in order to which multiple specific embodiments disclosed herein are explicitly described, lists perhaps
More details.But those skilled in the art will be understood that not include all tools being discussed below
The invention currently advocated is practiced in the case where body details.In other examples, in order not to obscure the invention, well-known zero
Part is not described.
In bion action selection and enhancing study combination in any environment successful adaptation and sprawling be
It is very important.This is also to be applicable in for the successful operating of intelligent agent.Here what is presented is for by neuromorphic
Creative action selection/enhancing network biologically of reason device control agent designs and implements.
The modeling process embodied can be described as the coupling of calculation biology and engineering science.In history, implement artificial
The strategy of intelligence fails the factor for leading to have real emergent properties.Therefore, dispose machine individual human and it is desirable that its from environment
It is still unadvisable for learning the behavior of bion.Equally, neural model needs complicated and multiplicity input signal
To accurately replicate intravital activity.A kind of method for creating these complex stimulus is by the way that model is immersed in energy
In enough true or virtual environment that feedback is provided.
Conceptually, action selection is to compete the arbitration of signal.In mammalian nervous system, by believing in multiple inputs
Properly selecting between number, when information that the complicated circuit of basal ganglion flows in opening volume cortex are effective.It should
Selection mechanism can be to simple action until complex behavior and understanding processing have an impact.Although exceedingly simplifying, it has
Help BG is associated with circuit multiplexer, according to current system mode, actively connection inputs the circuit multiplexer
And output.
Enhancing or award study (RL) are the positive result maximumlly enhancings to action or decision for selecting these.This
Similar to tool conditional reflection, wherein stimuli responsive test causes the response of award enhanced, and the response to decay not by
Enhancing.Enhancing study in neural network is the ideal substitution to supervised learning algorithm.Supervised learning needs intelligent guidance signal,
The intelligent guidance signal needs to have detailed understanding to task, and the exploitation for enhancing study can not need any independently of task
First standby knowledge.Only need the quality of the output signal as the response to input signal and the current ambient conditions of network.
One in accordance with an embodiment of the present disclosure in, can by band leak integral triggering (LIF) model in neural network
Neural Meta Model.LIF model is defined by equation 1.
Wherein
Cm is membrane capacitance,
I be foreign current and synaptic currents and,
Gleak is the specific conductance of leak channel,
Erest is the reversal potential of certain kinds cynapse.
As the electric current of input model neuron increases, membrane voltage is also scaled up, until reaching threshold voltage.At this time
Action potential is triggered, and membrane voltage resets to stand-by value.Neuron models are in 2 milliseconds of high resistant (refractory) period, this
When membrane voltage do not allow that any variation occurs.If electric current is removed before reaching threshold value, voltage reduces to Erest.LIF mould
Type is one of the smallest neural model of computational intensity, and still is able to many aspects of duplication neurobehavioral.
Connection between neuron or cynapse is modeled by the cynapse based on specific conductance.The general type of the influence
As equation 2 defines.
gsyn=gmax·geff·(V-Esyn). (2)
Wherein
Gmax is the maximum conductance coefficient of certain kinds cynapse,
Geff is the current synaptic efficacy between [0, geffmax],
Esyn is the reversal potential of certain kinds cynapse.
For the buffering and re-absorption of analog neuron transmitting element, can will be taken action in the presynaptic current potential based on defined time constant
The influence decaying on neuron having.Equation 3 can be used to extract the process.
Study on synaptic levels is to rely on plasticity rule by peak hour to realize, such as Song, S.,
K.D.Miller and L.F.Abbott (2000) is at 919-926 pages of Nature neuroscience (9) of 2000
“Competitive Hebbian Learning through Spike-timing Dependent Synaptic
Peak hour described in Plasticity " relies on plasticity rule, as shown in equation 4.
geff→geff+geffmaxF(Δ t) (4)
Wherein,
Δ t=tpre-tpost
If (geff< 0), then geff→0
If (g > geffmax), then geff→geffmax。
The global parameter value that can be used in one embodiment is presented in table 1.Using Euler's integral with 1 millisecond (ms) when
Between step-length to governing equation carry out digital integration.
Table 1: world model's parameter
Fig. 1-3 illustrates three different neural network embodiments.Initially, each of these neural networks is to them
Environment there is no any understanding or inherent understand.Its performance is the feedback by award or punishment signal form from environment
It practises, award or punishment signal are compiled as random or structured furcella event.Mind is reinforced or weakened to these signals
Through the Synaptic junction between member；Strengthen action appropriate.
First model shown in Fig. 1 is pure feedforward network, and the feedforward network is by being configured in N number of channel
Whole excitatory neuron compositions.The neural network of Fig. 1 has N number of channel.Each of N number of channel channel all has one group
Input 12, one groups of neuron, 14, one groups of output neuron award neuron 16.
In one embodiment, each group input neuron 12 is connected to equal specific conductance all with equal probability
Each group output neuron 14, so that it is guaranteed that specific input-output is to there is no inherent biass.In another embodiment,
Each group input neuron 12 is randomly connected to each group output neuron 14.Large-scale reality of the embodiment for these networks
It applies scheme and is passed to limitation as caused by neuromorphic system and be even more important.
The input neuron 12 in each channel is connected to the output neuron 14 in each channel by cynapse 18.It is in table 2
One group of parameter of the model that can be used for Fig. 1 is showed.The Synaptic junction 18 inputted between neuron 12 and output neuron 14 is logical
It crosses and fully enters what neural tuple 12 created at random, so that it is guaranteed that bias is not present between input and output channel.
Award neuron 16, which is received, can pass through the input of environment sensing from environment input 20.The award mind in each channel
The output neuron 20 of respective channel is coupled only to by cynapse 22 through member.If channel environment input be it is positive, accordingly
The output neuron 14 in channel is awarded and has the response of enhancing.If the environment input in channel is negative, respective channel
Output neuron 14 it is penalized and have decaying response.
Can be defined by equation 1 band leakage integral triggering (LIF) model to input neuron 12, output neuron 14,
Neuron 16 is awarded to model.It can cynapse 18 and 22 be built by peak hour dependence plasticity model (STDP) by equation 4
Mould.
Table 2: the parameter of network is stimulated
A. neuron parameter
B. it connects
Fig. 2 illustrates another neural network according to the disclosure, has lateral inhibition between the output group of the neural network.
The neural network of Fig. 2 creates closed center around network, wherein most of positive groups inhibit other output groups.This is not only one
A biologically more feasible network, and it also provides more controls to selection process.One group for the model
Parameter can be parameter shown in table 3.One main aspect of neural network is to inhibit the diffusion connection of neuron 36.Each
One group of inhibition neuron 36 in channel all project in addition to belonging to the group inhibits neuron 36 each of other than channel it is other logical
The output neuron 32 in road.
The neuroid of Fig. 2 has N number of channel.Each channel in this N number of channel has one group of input neuron 30, one
Group 32, one groups of output neuron, 34, one groups of award neuron inhibition neuron 36.One group of input neuron 30 in each channel
One group of output neuron 32 in each channel is connected to by cynapse 38.
In one embodiment, these group input neurons 30 are connected to institute with equal specific conductance with equal probability
There is the output neuron 32 of group, so that it is guaranteed that specific input-output is to there is no inherent biass.In another embodiment,
Each group input the Synaptic junction 38 between neuron 30 and each group output neuron 32 by whole input nerve tuples 30 with
Connect to machine.
The award neuron 34 in each channel, which is received, can pass through the input of environment sensing from environment input 40.It is each logical
The award neuron 34 in road is coupled only to the output neuron 32 of respective channel by cynapse 42.If the environment input in channel is
Positive, then the output neuron 32 of respective channel is awarded and has the response of enhancing.If channel environment input be it is negative,
Then the output neuron 32 of respective channel is penalized and has the response of decaying.
The output neuron 32 in each channel is connected to the inhibition neuron 36 of respective channel by cynapse 46.In channel
Neuron 36 is inhibited to be couple to the output neuron 32 in each other channels by cynapse 44, but the inhibition nerve in channel
Member 36 is not coupled to the output neuron 32 for inhibiting channel belonging to neuron 36.
When the response from the output neuron 32 for inhibiting channel belonging to neuron 36 increases, inhibit neuron 36 can
Inhibit the response of the output neuron 32 in each other channels by cynapse 44.
Can be defined by equation 1 band leakage integral triggering (LIF) model to input neuron 30, output neuron 32,
It awards neuron 34, neuron 36 is inhibited to model.The right by peak hour dependence plasticity model (STDP) of equation 4 can be passed through
Cynapse 38,42,44,46 models.
Table 3: the parameter of lateral inhibition network
A. neuron parameter
B. it connects
Fig. 3 illustrates basal ganglion (BG) neural network according to the disclosure.The neuron network simulation of Fig. 3 BG
The physiological activity of direct path, wherein Substantia nigra reticulata (SNr) neuron 56 is tatanic movable, and is put with 30Hz or so
Electricity.Black substance is a part of basal ganglion, and reticular part is a part of black substance.The Basal Activity of SNr neuron 56 is by line
The incoming control of the inhibition of shape somatic nerves member 52, which results in the mechanism of disinthibiting of action.Study occurs in cortical neuron 50
Between striatal neuron 52, to develop input-output channel combination appropriate.One group is illustrated in table 4 can be used for
The parameter of the model.
Table 4: the parameter of basal ganglion direct path
A. neuron parameter
B. it connects
Physiologically SNr neuron 56 is tatanic movable.But this is not spontaneous for reproducible for the LIF neuron of equation 1
Sexuality.In order to make up, Poisson random stimulus input 68 is injected into SNr neuron 56.In addition, can by low-level uniformly with
Machine noise is injected into network.
The neural network of Fig. 3 has N number of channel.Each channel in this N number of channel has one group 50, one groups of cortical neuron
52, one groups of striatal neuron, 54, one groups of award neuron SNr neuron 56.The cortical neuron 50 in each channel passes through prominent
Touching 58 is connected to each striatal neuron channel.
In one embodiment, each group cortical neuron 50 is connected to equal specific conductance all with equal probability
Each group striatal neuron 52, so that it is guaranteed that specific cortical/striatal is to there is no inherent biass.In another embodiment
In, each group cortical neuron 50 is randomly connected to each group striatal neuron 52.
One group of striatal neuron 52 in one channel is connected to one group of line in each other channels by cynapse 60
Shape somatic nerves member 52.
Award neuron 54, which is received, can pass through the input of environment sensing from environment input 62.The award mind in each channel
The striatal neuron 52 of respective channel is coupled only to by cynapse 64 through member 54, award neuron 54 is striatal neuron 52
A part.If channel environment input be it is positive, the striatal neuron 52 of respective channel award and have enhance
Response.If channel environment input be it is negative, the striatal neuron 52 of respective channel it is penalized and have decaying
Response.
The striatal neuron 52 in each channel is connected solely to the SNr neuron 56 of respective channel by cynapse 66.Poisson
Random stimulus input 68 is injected into the SNr neuron 56 in each channel.
Band leakage integral triggering (LIF) model that can be defined by equation 1 is to cortical neuron 50, striatal neuron
52, neuron 54 is awarded, SNr neuron 56 models.Plasticity model (STDP) can be relied on by peak hour by equation 4
Cynapse 58,60,64,66 is modeled.
Study in these networks is to be injected to drive by conditional stimulus.Changeless spiking can be sent to defeated
Enter group and all award groups.The time point of the signal of destination channel is postponed, so that input group and desired output group
Between the enhanced while all other channel of cynapse study be suppressed.The time point of these signals depends in equation 4 selecting
The value selected.Punishment signal can be injected by removing delay from target award group and inhibiting the activity of other output groups.
This is a kind of side of only architecture for developing the network that these are used to create any input/output combination
Formula.Synapse weight can be realized using any conspicuous cloth, actor-commentator, award-modulation or tip-award learning rules
Identical modulation.
Equally, LIF neuron is an example of workable neuron models.It is any multiple signals to be carried out
Integral operation and the mathematical model for being converted into discrete time-event can be used in these networks.
Finally, specific connectivity is not conclusive to performance.The quantity for increasing the connection in each cell can mention
High stability and plasticity.
The model of Fig. 1 be the neuromorphic processor original based on memristor constraint condition under realize.Fig. 4 A-
4C shows an exemplary enhancing study script.The activity rate that Fig. 4 A illustrates the exemplary script is drawn.Activity be using
What mobile Gauss weighted window calculated.Fig. 4 B shows defeated enrolled spike grating.Fig. 4 C shows the spike grating of output group.
Each stage by the center Fig. 4 A alphabetic flag.Fig. 5 A-5F respectively show 0 second, 10 seconds, 11 seconds, 21 seconds, 22 seconds,
Synapse weight at 33 seconds.
In stage A, network is initialized with all input/output connection all cynapse use values with 0.25；
As in Fig. 5 A shown by the thermal map of the average weight between input/output group.
In stage B, injection Poisson stochastic inputs for 10 seconds are to establish the Basal Activity of network in continuous passage.Figure
Average synapse weight matrix caused by being illustrated in 5B.
In stage C, alternate reward signal is sent to create single input/output pair.Weight matrix is now by scheming
Diagonal line shown in 5C dominates.
In stage D, injection for 10 seconds and the duplicate Poisson input signal of stage B above.Hereafter, shown in Fig. 5 D
Weight matrix show created input/output pair further enhance and the lasting inhibition of other connections.
In stage E, a set of opposite input/output is established using alternately reward signal and is associated with.In order to be carried out to network
Stable retraining, award agreement need be original training two double-lengths.New weight matrix is illustrated in Fig. 5 E.
In stage F, duplicate Poisson input in 10 seconds shows the newly created input/output pair in Fig. 5 F.
In order to show lateral inhibition network, a kind of virtual environment of ball-type of rattling is implemented.Fig. 6 is the model of the environment.Trip
The position of ball 70 in play space is sent to the neural channel of many discretizations.Each of these channels all substantial generation
One vertical column of table cribbage-board.These inputs are that have the Poisson of ratio defined in Gaussian curve described below sharp at random
Peak event.This improves noise inputs to the lap between channel.By the mechanism of Winner-take-all, network sends racket 72
Position signal.
Initially, network does not have any understanding or the understanding of inherence for how to carry out the game.By random to be compiled as
The feedback that the award of spike event and punishment signal provide learns operating condition.The stimulation for entering network is by ball 70
The position of each spatial channel is determined.The signal strength of each spatial channel is by the Gauss based on channel position
Function is sampled to calculate.The position of ball 70 in drawing determines peak amplitude and the center of Gaussian function, the Gauss
Function is defined as follows:
fxc(X*)=ac- ((xc-x*)2/2c2)
Wherein
A is the peak amplitude of this function
B is the center of this function
C is the space width of Gaussian function
Xc is the zero dimension position in channel
Peak amplitude and Gauss center are defined as follows:
A=Y*·Rmax (2)
b=X* (3)
Wherein
Y* is zero dimension position of the ball in y-dimension,
Rmax is the maximum input stimulus in spike per second,
X* is zero dimension position of the ball in x dimension.
This is shown in Fig. 7, and wherein space width c is 0.05.The generation when ball 70 reaches the bottom of cribbage-board 74
Award and punishment to network.The example sexual stimulus that Fig. 7 A illustrates two spaces channel is drawn.Fig. 7 B illustrates two companies
Stimulation overlapping between continuous spatial channel.Fig. 7 C illustrates the example sexual stimulus of the different location of ball 70.
Fig. 8 and Fig. 9 is illustrated when space width c is 0.025 in 0-25 second of Fig. 8 A, the 50-75 second of Fig. 8 B, Fig. 8 C
125-150 seconds results.Figure 10 illustrates the overall accuracy for the model that space width c is 0.025.
The neural network of Fig. 1-3 can be implemented by passive and active electronic building brick, including transistor, resistor, capacitor.Institute
Stating neural network can also be implemented by computer or processor.The processor of type workable for a kind of is the nerve based on memristor
Morphological process device.
Thus according to the requirement of patent statute, invention has been described, it will be appreciated by those skilled in the art that how
Modifications and changes are made to the present invention so that it meets specific requirement and condition.Present invention disclosed herein can not departed from
Scope and spirit in the case where these modifications and changes are made to the present invention.
Presented above is according to laws and regulations requirement to this hair exemplary and detailed description of preferred embodiment purpose
It is bright to be illustrated and openly.Its purpose is not exhaustive or limits the invention in described (multiple) concrete form, and only uses
In make others skilled in the art be understood that the present invention how to be adapted to specific purposes or embodiment.To this field
It obtains employment for technical staff, modifies and be obvious a possibility that change.To including deviation, part size, specific
The description of the exemplary embodiment of operating condition, engineering specification etc. is not intended to limit, can various embodiments it
Between change or with state of art change and be changed, not imply any restrictions.Applicant is directed to the prior art
It is made that the disclosure, but still expected further improvement, and can be by considering that these improve (i.e. following " the existing skill
Art ") present invention will be stood good in future.It is intended that the scope of the present invention is by the claims hereof and its can
Applicable equivalent is limited.Unless explicitly stated otherwise, the singular being related to otherwise in claims is not meant as " one
A and only one ".In addition, for any element of the invention, component or method, process steps, regardless of these elements, component,
Or whether step is distinctly claimed in detail in the claims, they are meant to gratuitously contribute to the public.Of the invention
Element, otherwise cannot be Section 112 of volume 35 according to United States Code No. unless clearly quoted using phrase " tool being used for ... "
6th section is explained the element of claim, and step herein is using " including the steps that ... " unless clearly quoted,
Otherwise the method for claim or process steps can not be explained according to aforesaid clause.
Following design is at least disclosed herein:
Conceive 1. a kind of for enhancing the neural model of study and selection of taking action, which includes:
Multiple channels；
One group of input neuron in each channel；
One group of output neuron in each channel, every group of input neuron in each channel are couple to each logical
Every group of output neuron in road；
One group of award neuron in each channel, wherein the reception of every group of award neuron inputted from environment it is defeated
Enter, and wherein the award neuron in each channel is coupled only to the output neuron in channel belonging to award neuron；
Wherein, if channel environment input be it is positive, the output neuron of respective channel award and have enhance
Response；
Wherein, if channel environment input be it is negative, the output neuron of respective channel it is penalized and have decaying
Response.
Neural model of the design 2. as described in design 1, wherein every group of output neuron in each channel passes through with table
The cynapse that now following peak hour relies on plasticity is couple to every group of input neuron in each channel:
geff→geff+geffmaxF(Δt)
Wherein,
Δ t=tpre-tpost
If (geff< 0), then geff→0
If (g > geffmax), then geff→geffmax。
Neural model of the design 3. as described in design 1, wherein every group of input neuron, every group of output neuron, every group of prize
Appreciating neuron is to trigger (LIF) model modeling by showing following integrating with leakage:
Wherein
Cm is membrane capacitance,
I be external and synaptic currents and,
Gleak is the specific conductance of leak channel,
Erest is the reversal potential of certain kinds cynapse.
Design 4. as design 1 as described in neural model, wherein each group input neuron with equal probability with equal electricity
It leads coefficient and is connected to all each group output neurons.
Neural model of the design 5. as described in design 1, wherein each group input neuron is randomly to be connected to each group output
Neuron.
Neural model of the design 6. as described in design 1, wherein the neural model is by the neuromorphic based on memristor
Device is managed to realize.
Conceive 7. a kind of for enhancing the neural model of study and selection of taking action, which includes:
Multiple channels；
Input nerve tuple in each channel；
Output nerve tuple in each channel, every group of input neuron in each channel are couple to each channel
In every group of output neuron；
Award nerve tuple in each channel, wherein the reception of every group of award neuron inputted from environment it is defeated
Enter, and wherein the award neuron in each channel is coupled only to the output neuron in channel belonging to award neuron；
Inhibition nerve tuple in each channel, wherein every group of inhibition neuron receives and inhibit neuron from the group
The input of output nerve tuple in affiliated same channel, and the inhibition nerve tuple in one of channel has in place
In in addition to the output of the output neuron belonging to the group inhibits neuron each of other than channel in other channels；
Wherein, if the input of tuple neural for the award in a channel is positive, the output nerve of respective channel
Tuple is awarded and has the response of enhancing；And
Wherein, if the input of tuple neural for the award in a channel is negative, the output nerve of respective channel
Tuple is penalized and has the response of decaying.
Neural model of the design 8. as described in design 7, in which:
Wherein every group of output neuron in each channel is couple to by the cynapse for relying on plasticity with peak hour
Every group of input neuron in each channel；
Each award neuron is couple to output neuron by the cynapse for relying on plasticity with peak hour；
For the output nerve tuple of every group of inhibition neuron inhibited in one channel of phase belonging to neuron from the group
Input be by with peak hour rely on plasticity synaptic input；And
The output of every group of inhibition neuron in one channel is coupled by relying on the cynapse of plasticity with peak hour
To in addition to the output neuron in other each channels belonging to the group inhibits neuron other than channel；
Wherein the peak hour of each cynapse, it is as follows to rely on plastic sex expression:
geff→geff+geffmaxF(Δt)
Wherein,
Δ t=tpre-tpost
If (geff< 0), then geff→0
If (g > geffmax), then geff→geffmax。
Neural model of the design 9. as described in design 7, wherein every group of input neuron, every group of output neuron, every group of prize
Appreciate neuron, every group of inhibition neuron is to trigger (LIF) model modeling by showing following integrating with leakage:
Wherein
Cm is membrane capacitance,
I be external and synaptic currents and,
Gleak is the specific conductance of leak channel,
Erest is the reversal potential of certain kinds cynapse.
Design 10. as design 7 as described in neural model, wherein each group input neuron with equal probability with equal electricity
It leads coefficient and is connected to all each group output neurons.
Neural model of the design 11. as described in design 7, wherein each group input neuron is randomly to be connected to each group output
Neuron.
Neural model of the design 12. as described in conceiving 7, wherein when channel belonging to one group of inhibition neuron
When the response of output neuron increases, the sound for each group output neuron for inhibiting neuron to inhibit in each other channels
It answers.
Neural model of the design 13. as described in design 7, wherein the neural model is by the neuromorphic based on memristor
Device is managed to realize.
Conceive a kind of 14. basal ganglion neural network models comprising:
Multiple channels；
Cortical neurons tuple in each channel；
Striatal neuron group in each channel, every group of striatal neuron in each channel are couple to each
Every group of cortical neuron in channel；
Award nerve tuple in each channel, wherein the reception of every group of award neuron inputted from environment it is defeated
Enter, and wherein the award neuron in each channel is coupled only to the striatal neuron in channel belonging to award neuron；
And
The neural tuple of Substantia nigra reticulata (SNr) in each channel, wherein every group of SNr neuron is coupled only to SNr
Striatal neuron group in channel belonging to neuron；
Wherein, if the input of tuple neural for the award in a channel is positive, the corpus straitum mind of respective channel
The response of enhancing is awarded and had through tuple；
Wherein, if the input of tuple neural for the award in a channel is negative, the corpus straitum mind of respective channel
Response penalized through tuple and that there is decaying；
Wherein every group of SNr neuron is tatanic movable and is corpus straitum in the channel as belonging to SNr neuron
The incoming inhibition of the inhibition of neuron.
Basal ganglion neural network model of the design 15. as described in design 14, in which:
Every group of cortical neuron in each channel is couple to each by the cynapse for relying on plasticity with peak hour
Every group of striatal neuron in channel；
Every group of striatal neuron in each channel is couple to often by the cynapse for relying on plasticity with peak hour
Striatal neuron in a other channels；
The award neuron in each channel is couple in same channel by the cynapse for relying on plasticity with peak hour
Striatal neuron group；
Every group of SNr neuron is couple to belonging to this group of SNr neuron by the cynapse for relying on plasticity with peak hour
Same channel in striatal neuron group；And
Wherein the peak hour of each cynapse, it is as follows to rely on plastic sex expression:
geff→geff+geffmaxF(Δt)
Wherein,
Δ t=tpre-tpost
If (geff< 0), then geff→0
If (g > geffmax), then geff→geffmax。
Basal ganglion neural network model of the design 16. as described in design 14, wherein every group of cortical neuron, every group of line
Shape somatic nerves member, every group of award neuron, every group of SNr neuron are by showing following band leakage integral triggering (LIF) model
Modeling:
Wherein
Cm is membrane capacitance,
I be external and synaptic currents and,
Gleak is the specific conductance of leak channel,
Erest is the reversal potential of certain kinds cynapse.
Basal ganglion neural network model of the design 17. as described in design 14, wherein each group cortical neuron is with equal
Probability and equal specific conductance be connected to all each group striatal neurons.
Basal ganglion neural network model of the design 18. as described in design 14, wherein each group cortical neuron is random
Ground is connected to each group striatal neuron.
Basal ganglion neural network model of the design 19. as described in design 14, wherein Poisson random stimulus is injected into
In each group SNr neuron.
Basal ganglion neural network model of the design 20. as described in design 14, wherein uniform random noise is injected into
In each group SNr neuron.
Basal ganglion neural network model of the design 21. as described in design 14, wherein the basal ganglion nerve net
Network model is realized by the neuromorphic processor based on memristor.
Claims (33)
1. a kind of for enhancing the neural network of study and selection of taking action, which includes:
Multiple channels；
Input nerve tuple in each channel, wherein input nerve tuple includes passive electric components and active electricity
Sub-component, the passive electric components and the active electronic component include transistor, resistor and capacitor；
Output nerve tuple in each channel, each input nerve tuple in each channel are couple in each channel
Each output nerve tuple, wherein the output neuron group includes passive electric components and active electronic component, this is passive
Electronic building brick and the active electronic component include transistor, resistor and capacitor；With
Award nerve tuple in each channel, wherein each input awarding neural tuple reception and being inputted from environment,
And wherein the award neuron in each channel is coupled only to the output neuron in channel belonging to award neuron, and its
Described in award neuron include passive electric components and active electronic component, the passive electric components and the active electronic component
Including transistor, resistor and capacitor；
Wherein, when the input of the environment in channel is positive, the output neuron of respective channel is awarded and has the response of enhancing；
Wherein, when the input of the environment in channel is negative, the output neuron of respective channel is penalized and has the response of decaying.
2. neural network as described in claim 1, wherein each output nerve tuple in each channel is as follows by having
The cynapse that the peak hour of performance relies on plasticity is couple to each input nerve tuple in each channel:
geff→geff+geffmaxF(Δt)
Wherein,
Δ t=tpre-tpost
If (geff< 0), then geff→ 0,
If (g > geffmax), then geff→geffmax。
3. neural network as described in claim 1, wherein it is each input neural tuple with equal probability with equal conductance system
Number is connected to all each output nerve tuples.
4. neural network as described in claim 1, wherein each nerve tuple that inputs randomly is connected to each output nerve tuple.
5. neural network as described in claim 1, wherein each input nerve tuple in each channel passes through the first cynapse
The each output nerve tuple being couple in each channel, and wherein the award neuron in each channel passes through the second cynapse
The output neuron being couple in channel belonging to award neuron.
6. neural network as claimed in claim 5, wherein each input nerve tuple, each output nerve tuple and each prize
Appreciating neural tuple has following band leakage integral triggering (LIF) behavior:
Wherein
Cm is membrane capacitance,
I be foreign current and synaptic currents and,
Gleak is the specific conductance of leak channel,
Erest is reversal potential.
7. neural network as claimed in claim 5, wherein the first cynapse and the second cynapse have the spike for meeting following equation
Time Dependent plasticity:
gsyn=gmax·geff·(V-Esyn)
Wherein,
gmaxIt is the maximum conductance coefficient of the first cynapse and the second cynapse,
geffIt is 0 and geffmaxBetween current synaptic efficacy, and
EsynIt is the reversal potential of the first cynapse and the second cynapse.
8. neural network as claimed in claim 7, in which:
geff→geff+geffmaxF(Δt)
Wherein,
Δ t=tpre-tpost
If (geff< 0), then geff→ 0,
If (geff> geffmax), then geff→geffmax。
9. neural network as described in claim 1, wherein the neural network includes the neuromorphic processing based on memristor
Device.
10. neural network as claimed in claim 6, in which:
Presynaptic takes action current potential to each influence for inputting neural tuple, each output nerve tuple and the neural tuple of each award
Meet
11. a kind of for enhancing the neural network of study and selection of taking action, which includes:
Multiple channels；
Input nerve tuple in each channel, wherein input nerve tuple includes passive electric components and active electricity
Sub-component, the passive electric components and the active electronic component include transistor, resistor and capacitor；
Output nerve tuple in each channel, each input nerve tuple in each channel are couple in each channel
Each output nerve tuple, wherein the output neuron group includes passive electric components and active electronic component, this is passive
Electronic building brick and the active electronic component include transistor, resistor and capacitor；
Award nerve tuple in each channel, wherein each input awarding neural tuple reception and being inputted from environment,
And wherein the award neuron in each channel is coupled only to the output neuron in channel belonging to award neuron, and its
Described in award neuron include passive electric components and active electronic component, the passive electric components and the active electronic component
Including transistor, resistor and capacitor；With
Inhibition nerve tuple in each channel, wherein each inhibition nerve tuple receives and comes from the inhibition nerve tuple institute
The input of the output nerve tuple in same channel belonged to, and the inhibition nerve tuple in one of channel has to being located at
The output of output neuron each of other than the affiliated channel of inhibition nerve tuple in other channels, and it is wherein described
Inhibiting neural tuple includes passive electric components and active electronic component, and the passive electric components and the active electronic component include
Transistor, resistor and capacitor；
Wherein, when the environment input of the neural tuple of the award for a channel is positive, the output neuron of respective channel
Group is awarded and has the response of enhancing；
Wherein, when the environment input of the neural tuple of the award for a channel is negative, the output neuron of respective channel
Group is penalized and has the response of decaying.
12. neural network as claimed in claim 11, in which:
Wherein each output nerve tuple in each channel is couple to often by the cynapse for relying on plasticity with peak hour
Each input nerve tuple in a channel；
Award neuron in each channel is couple to output neuron by the cynapse for relying on plasticity with peak hour；
The input of output nerve tuple in each same channel belonging to the inhibition nerve tuple for inhibiting neural tuple
It is to be provided by the cynapse for relying on plasticity with peak hour；And
The output of each inhibition nerve tuple in one channel is couple to by the cynapse for relying on plasticity with peak hour
Output neuron each of other than the affiliated channel of inhibition nerve tuple in other channels；
Wherein the peak hour of each cynapse, it is as follows to rely on plastic sex expression:
geff→geff+geffmaxF(Δt)
Wherein,
Δ t=tpre-tpost
If (geff< 0), then geff→ 0,
If (g > geffmax), then geff→geffmax。
13. neural network as claimed in claim 11, wherein it is each input neural tuple with equal probability with equal conductance
Coefficient is connected to all each output nerve tuples.
14. neural network as claimed in claim 11, wherein each nerve tuple that inputs randomly is connected to each output neuron
Group.
15. neural network as claimed in claim 11, wherein when from the output nerve for inhibiting channel belonging to neural tuple
When the response of member increases, the response for each output nerve tuple for inhibiting neuron to inhibit in each other channels.
16. neural network as claimed in claim 11,
Wherein each input nerve tuple in each channel is couple to the mind of each output in each channel by the first cynapse
Through tuple,
Wherein the award neuron in each channel is coupled only to the output in channel belonging to award neuron by the second cynapse
Neuron,
Wherein the neural tuple of each inhibition is received by third cynapse in the same channel belonging to the inhibition nerve tuple
Output nerve tuple input, and
The neural tuple of inhibition in one of channel has the output coupled by the 4th cynapse, and the output, which is couple to, to remove
Output neuron each of except the affiliated channel of inhibition nerve tuple in other channels.
17. neural network as claimed in claim 16, wherein each input nerve tuple, each output nerve tuple, each
Awarding neural tuple and the neural tuple of each inhibition, there is following band to leak integral triggering (LIF) behavior:
Wherein
Cm is membrane capacitance,
I be foreign current and synaptic currents and,
Gleak is the specific conductance of leak channel,
Erest is reversal potential.
18. neural network as claimed in claim 16, wherein the first cynapse, the second cynapse, third cynapse and the 4th cynapse tool
There is the peak hour for meeting following equation to rely on plasticity:
gsyn=gmax·geff·(V-Esyn)
Wherein,
gmaxIt is the maximum conductance coefficient of the first cynapse and the second cynapse,
geffIt is 0 and geffmaxBetween current synaptic efficacy, and
EsynIt is the reversal potential of the first cynapse and the second cynapse.
19. neural network as claimed in claim 18, in which:
geff→geff+geffmaxF(Δt)
Wherein,
Δ t=tpre-tpost
If (geff< 0), then geff→ 0,
If (geff> geffmax), then geff→geffmax。
20. neural network as claimed in claim 11, wherein the neural network includes at the neuromorphic based on memristor
Manage device.
21. neural network as claimed in claim 17, in which:
Presynaptic takes action current potential to the neural tuple of each input, each output nerve tuple, each award nerve tuple and each
The influence of neural tuple is inhibited to meet
22. a kind of basal ganglion neural network comprising:
Multiple channels；
Cortical neurons tuple in each channel, wherein the cortical neuron group includes passive electric components and active electricity
Sub-component, the passive electric components and the active electronic component include transistor, resistor and capacitor；
Striatal neuron group in each channel, each striatal neuron group in each channel are couple to each logical
Each Cortical neurons tuple in road, wherein the striatal neuron group includes passive electric components and active electronic component,
The passive electric components and the active electronic component include transistor, resistor and capacitor；
Award nerve tuple in each channel, wherein each input awarding neural tuple reception and being inputted from environment,
And wherein the award neuron in each channel is coupled only to the striatal neuron in channel belonging to award neuron, and
Wherein the neural tuple of the award includes passive electric components and active electronic component, the passive electric components and the active electronic
Component includes transistor, resistor and capacitor；With
Substantia nigra reticulata nerve tuple in each channel, wherein each Substantia nigra reticulata nerve tuple is coupled only to black substance
Striatal neuron group in channel belonging to reticular part neuron, and wherein the Substantia nigra reticulata nerve tuple includes nothing
Source electronic building brick and active electronic component, the passive electric components and the active electronic component include transistor, resistor and electricity
Hold；
Wherein, when the environment input of the neural tuple of the award for a channel is positive, the corpus straitum nerve of respective channel
Tuple is awarded and has the response of enhancing；
Wherein, when the environment input of the neural tuple of the award for a channel is negative, the corpus straitum nerve of respective channel
Tuple is penalized and has the response of decaying；
Wherein each Substantia nigra reticulata nerve tuple is tatanic movable and the channel as belonging to Substantia nigra reticulata neuron
In the inhibition of striatal neuron incoming inhibit.
23. basal ganglion neural network as claimed in claim 22, in which:
Each Cortical neurons tuple in each channel is couple to each logical by the cynapse for relying on plasticity with peak hour
Each striatal neuron group in road；
Each striatal neuron group in each channel is couple to each by the cynapse for relying on plasticity with peak hour
Striatal neuron in other channels；
The channel of each award neuron is couple to one in same channels by the cynapse for relying on plasticity with peak hour
Group striatal neuron；
Each Substantia nigra reticulata nerve tuple is couple to the Substantia nigra reticulata by the cynapse for relying on plasticity with peak hour
Striatal neuron group in same channel belonging to neural tuple；And
Wherein the peak hour of each cynapse, it is as follows to rely on plastic sex expression:
geff→geff+geffmaxF(Δt)
Wherein,
Δ t=tpre-tpost
If (geff< 0), then geff→ 0,
If (g > geffmax), then geff→geffmax。
24. basal ganglion neural network as claimed in claim 22, wherein each Cortical neurons tuple is with equal probability
All each striatal neuron groups are connected to equal specific conductance.
25. basal ganglion neural network as claimed in claim 22, wherein each Cortical neurons tuple is randomly connected to
Each striatal neuron group.
26. basal ganglion neural network as claimed in claim 22, wherein Poisson random stimulus is injected into each black substance
In reticular part nerve tuple.
27. basal ganglion neural network as claimed in claim 22, wherein uniform random noise is injected into each black substance
In reticular part nerve tuple.
28. basal ganglion neural network as claimed in claim 22,
Wherein each striatal neuron group in each channel is couple to each cortex in each channel by the first cynapse
Neural tuple；
Wherein the award neuron in each channel is coupled only to the line shape in channel belonging to award neuron by the second cynapse
Somatic nerves member；And
Wherein each Substantia nigra reticulata nerve tuple is coupled only to channel belonging to Substantia nigra reticulata neuron by third cynapse
In striatal neuron group.
29. basal ganglion neural network as claimed in claim 28, wherein each Cortical neurons tuple, each corpus straitum are refreshing
There is following band leakage integral triggering (LIF) through tuple, each neural tuple and each Substantia nigra reticulata nerve tuple of awarding
Behavior:
Wherein
Cm is membrane capacitance,
I be foreign current and synaptic currents and,
Gleak is the specific conductance of leak channel,
Erest is reversal potential.
30. basal ganglion neural network as claimed in claim 28, wherein the first cynapse, the second cynapse and third cynapse tool
There is the peak hour for meeting following equation to rely on plasticity:
gsyn=gmax·geff·(V-Esyn)
Wherein,
gmaxIt is the maximum conductance coefficient of the first cynapse and the second cynapse,
geffIt is 0 and geffmaxBetween current synaptic efficacy, and
EsynIt is the reversal potential of the first cynapse and the second cynapse.
31. basal ganglion neural network as claimed in claim 30, in which:
geff→geff+gefmaxF(Δt)
Wherein,
Δ t=tpre-tpost
If (geff< 0), then geff→ 0,
If (geff> geffmax), then geff→geffmax。
32. basal ganglion neural network as claimed in claim 22, wherein the basal ganglion neural network includes base
In the neuromorphic processor of memristor.
33. basal ganglion neural network as claimed in claim 29, in which:
Presynaptic takes action current potential to each Cortical neurons tuple, each striatal neuron group, the neural tuple of each award and often
The influence of a Substantia nigra reticulata nerve tuple meets
Applications Claiming Priority (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| US201261732590P | 2012-12-03 | 2012-12-03 | |
| US61/732,590 | 2012-12-03 | ||
| PCT/US2013/041451 WO2014088634A1 (en) | 2012-12-03 | 2013-05-16 | Neural model for reinforcement learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104823205A CN104823205A (en) | 2015-08-05 |
| CN104823205B true CN104823205B (en) | 2019-05-28 |
Family
ID=51896593
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201380063033.4A Expired - Fee Related CN104823205B (en) | 2012-12-03 | 2013-05-16 | For enhancing the neural model of study |
Country Status (2)
| Country | Link |
|---|---|
| EP (1) | EP2939187A4 (en) |
| CN (1) | CN104823205B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180174030A1 (en) * | 2016-12-15 | 2018-06-21 | Fu-Chang Hsu | Self-learning for neural network arrays |
| CN110428049B (en) * | 2019-08-21 | 2021-10-26 | 南京邮电大学 | Voltage type neural network based on polymorphic memristor and operation method thereof |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101432737A (en) * | 2004-07-15 | 2009-05-13 | 神经科学研究基金会 | Mobile brain-based device having a simulated nervous system based on the hippocampus |
| CN101971166A (en) * | 2008-03-14 | 2011-02-09 | 惠普开发有限公司 | Neuromorphic circuit |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| IL133384A0 (en) * | 1997-06-11 | 2001-04-30 | Univ Southern California | Dynamic synapse for signal processing in neural networks |
| US20090182697A1 (en) * | 2005-08-15 | 2009-07-16 | Massaquoi Steve G | Computer-Implemented Model of the Central Nervous System |
-
2013
- 2013-05-16 EP EP13860582.9A patent/EP2939187A4/en not_active Ceased
- 2013-05-16 CN CN201380063033.4A patent/CN104823205B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101432737A (en) * | 2004-07-15 | 2009-05-13 | 神经科学研究基金会 | Mobile brain-based device having a simulated nervous system based on the hippocampus |
| CN101971166A (en) * | 2008-03-14 | 2011-02-09 | 惠普开发有限公司 | Neuromorphic circuit |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104823205A (en) | 2015-08-05 |
| EP2939187A4 (en) | 2017-08-16 |
| EP2939187A1 (en) | 2015-11-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Gell-Mann | Complex adaptive systems | |
| Guo et al. | Attractivity analysis of memristor-based cellular neural networks with time-varying delays | |
| Milo et al. | Demonstration of hybrid CMOS/RRAM neural networks with spike time/rate-dependent plasticity | |
| KR101793011B1 (en) | Efficient hardware implementation of spiking networks | |
| US20150302296A1 (en) | Plastic action-selection networks for neuromorphic hardware | |
| Northrop | Introduction to complexity and complex systems | |
| CN101971166B (en) | Neuromorphic circuit | |
| Zheng et al. | Learning in memristor crossbar-based spiking neural networks through modulation of weight-dependent spike-timing-dependent plasticity | |
| US20150120627A1 (en) | Causal saliency time inference | |
| Carlson et al. | Biologically plausible models of homeostasis and STDP: stability and learning in spiking neural networks | |
| Wang et al. | Adaptive neuromorphic architecture (ANA) | |
| CN104823205B (en) | For enhancing the neural model of study | |
| Zhang et al. | Implementation of memristive neural networks with spike-rate-dependent plasticity synapses | |
| Orhan | The leaky integrate-and-fire neuron model | |
| US9418332B2 (en) | Post ghost plasticity | |
| Azghadi et al. | Design and implementation of BCM rule based on spike-timing dependent plasticity | |
| JP2021140320A (en) | Spiking neural network device and method for learning spiking neural network device | |
| Kochen | Advances in Cognitive Science: Steps Toward Convergence | |
| Sardi et al. | Dendritic learning as a paradigm shift in brain learning | |
| Yu et al. | CaMKII activation supports reward-based neural network optimization through Hamiltonian sampling | |
| James et al. | Guest editorial: special issue on large-scale memristive systems and neurochips for computational intelligence | |
| Ning et al. | Artificial neuron with somatic and axonal computation units: Mathematical and neuromorphic models of persistent firing neurons | |
| Buckley | A systemic analysis of the ideas immanent in neuromodulation | |
| Birkoben | Dynamics meets Morphology: towards Dymorph Computation | |
| Damper et al. | The Hi-NOON neural simulator and its applications |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190528 |