CN106875005B - Adaptive threshold neuronal messages processing method and system - Google Patents
Adaptive threshold neuronal messages processing method and system Download PDFInfo
- Publication number
- CN106875005B CN106875005B CN201710042067.5A CN201710042067A CN106875005B CN 106875005 B CN106875005 B CN 106875005B CN 201710042067 A CN201710042067 A CN 201710042067A CN 106875005 B CN106875005 B CN 106875005B
- Authority
- CN
- China
- Prior art keywords
- current
- threshold
- neuron
- adaptive threshold
- information
- 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.)
- Active
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
- G06N3/04—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
- 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
- G06N3/061—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using biological neurons, e.g. biological neurons connected to an integrated circuit
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Theoretical Computer Science (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Neurology (AREA)
- Microelectronics & Electronic Packaging (AREA)
- Feedback Control In General (AREA)
Abstract
The present invention relates to a kind of adaptive threshold neuronal messages processing method and system, which comprises receiving front-end spiking neuron output information;Read current PRF neuronal messages;According to the front pulse neuron output information and the current PRF neuronal messages, current PRF neuron output information is calculated;Current adaptive threshold variable and threshold potential are read, current adaptive threshold is calculated;When the current PRF neuron output information is greater than or equal to the adaptive threshold, export the current PRF neuron output information, and the current adaptive threshold variable according to the first adaptive threshold update model modification, otherwise the current PRF neuron output information, and the current adaptive threshold variable according to the second adaptive threshold update model modification are not exported then.The present invention can improve the information processing capability of impulsive neural networks with the granting frequency of each neuron in efficient balance whole network.
Description
Technical field
The present invention relates to artificial neural network technology fields, more particularly to adaptive threshold neuronal messages processing method
And system.
Background technique
The artificial neural network research overwhelming majority of today is still in von neumann machine software and high-performance of arranging in pairs or groups
It is realized in GPGPU (General Purpose Graphic Processing Units universal graphics processing unit) platform,
Hardware spending, energy consumption and the information processing rate of whole process all allow of no optimist.For this purpose, neuromorphic calculating field is fast in recent years
It hails exhibition, i.e., simulates the function of brain using hardware circuit direct construction neural network, it is intended to realize large-scale parallel, low
Energy consumption, the computing platform of sustainable complex patterns study.
However, in traditional neuromorphic system, how each spiking neuron in the entire neural network of efficient balance
Granting frequency so that each spiking neuron can play a role when handling information, be in traditional neural network urgently
It solves the problems, such as.
Summary of the invention
Based on this, it is necessary to for how the granting frequency of each spiking neuron in the entire neural network of efficient balance
The problem of, a kind of adaptive threshold neuronal messages processing method and system are provided, which comprises
Receiving front-end spiking neuron output information;
Read current PRF neuronal messages;
According to the front pulse neuron output information and the current PRF neuronal messages, current PRF mind is calculated
Through first output information;
Current adaptive threshold variable and threshold potential are read, and according to the current adaptive threshold variable and the threshold
It is worth current potential, calculates current adaptive threshold;
Judge whether the current PRF neuron output information is greater than or equal to the adaptive threshold, if so, defeated
The current PRF neuron output information out, and the current adaptive thresholding according to the first adaptive threshold update model modification
It is worth variable, if it is not,
It does not export the current PRF neuron output information then, and model modification institute is updated according to the second adaptive threshold
State current adaptive threshold variable.
The front pulse neuron output information includes: the output of front pulse neuron in one of the embodiments,
Pulse tip formation, front pulse neuron and current PRF neuron connection weight index;
The current PRF neuronal messages include: current time window width, pulse tip formation sequence in current time window
Column, history film potential information and film potential leakage information;
Then according to the front pulse neuron output information and the current PRF neuronal messages, current PRF is calculated
Neuron output information, comprising:
It is indexed according to the connection weight of the front pulse neuron and current PRF neuron, reads front pulse nerve
The connection weight of member and current PRF neuron;
According to pulse tip letter in the pulse tip formation and the current time window of front pulse neuron output
Sequence is ceased, pulse tip formation sequence in the current time window is updated, pulse tip formation in current time window is obtained and updates
Sequence;
According to pulse tip formation renewal sequence in the current time window width, the current time window, pass through decaying
Function calculates front pulse neuron and inputs information;
The company of information, the front pulse neuron and current PRF neuron is inputted according to the front pulse neuron
Weight, the history film potential information, film potential leakage information is connect to calculate current by spiking neuron computation model
Spiking neuron output information.
It is described when the current PRF neuron output information is more than or equal to described adaptive in one of the embodiments,
When answering threshold value, the current PRF neuron output information is exported, and according to the first adaptive threshold update model modification
Current adaptive threshold variable, further includes:
Determine that providing trigger flag information is to provide triggering, the granting trigger flag information includes providing triggering or providing
It does not trigger;
Refractory period timer is resetted, and updating the history film potential information is preset reset film potential information.
It is described in one of the embodiments, not export the current PRF neuron output information then, and according to second
Adaptive threshold updates current adaptive threshold variable described in model modification, further includes:
Determine that the granting trigger flag information does not trigger to provide;
Read the current time step of refractory period width and refractory period timer;
It is walked according to the current time of the refractory period width and the refractory period timer, judges current time whether not
It answers in the phase, if current time in the refractory period, by cumulative one time step of timing of the refractory period timer, does not update institute
State history film potential information;
If current time is not in the refractory period, by cumulative one time step of timing of the refractory period timer, and more
The new history film potential information is the current PRF neuron output information.
It is described in one of the embodiments, to read current adaptive threshold variable and threshold potential, and worked as according to described
Preceding adaptive threshold variable and the threshold potential, calculate current adaptive threshold, comprising:
Read random threshold value mask current potential, threshold bias, current adaptive threshold variable and random threshold value;
The random threshold value and the random threshold value mask current potential are subjected to step-by-step and operation, obtain threshold value random superposition
Amount;
According to the threshold value random superposition amount and the threshold bias, the threshold potential is determined;
According to the threshold potential and the current adaptive threshold variable, the current adaptive threshold is determined.
The current adaptive thresholding according to the first adaptive threshold update model modification in one of the embodiments,
It is worth variable, comprising:
It reads and provides threshold delta and the current adaptive threshold variable;
According to preset attenuation constant and the current adaptive threshold variable, first threshold is calculated;
The granting threshold delta is superimposed to the first threshold, obtains second threshold;
The current adaptive threshold variable is updated according to the second threshold.
The current adaptive thresholding according to the second adaptive threshold update model modification in one of the embodiments,
It is worth variable, comprising:
Read the current adaptive threshold variable;
According to the preset attenuation constant and the current adaptive threshold variable, third threshold value is calculated,
The current adaptive threshold variable is updated according to the third threshold value.
The output current PRF neuron output information in one of the embodiments, comprising:
It reading and provides enabled mark, the enabled mark of granting includes allowing to provide data or not allowing granting data, when
The granting enables to be identified as when allowing to provide data,
The granting trigger flag information is read, when the granting trigger flag information is to provide triggering;
Export the current PRF neuron output information.
Adaptive threshold neuronal messages processing method provided by the present invention, in current PRF neuron output information
In calculating, by reading current adaptive threshold variable and threshold potential, current adaptive threshold is calculated;According to current adaptive
Answer threshold value, it is determined whether output current PRF neuron output information, and determine the model for updating current adaptive dependent variable.This hair
The neuronal messages processing method of adaptive threshold provided by bright, can make the neuron threshold for currently having provided output information
Value increases, and provides difficulty next time and increases;And the neuron threshold value for not providing output information currently reduces, and provides difficulty drop next time
It is low, the granting frequency of each neuron in efficient balance whole network, so that each neuron can be sent out when handling information
The effect of waving greatly improves the information processing capability of impulsive neural networks.
According to the pulse tip formation of front pulse neuron output and described work as in one of the embodiments,
Pulse tip formation sequence in preceding time window updates pulse tip formation sequence in the current time window, obtains current time
Pulse tip formation renewal sequence in window, according to the current time window width, the front pulse neuron and current PRF
The connection weight of neuron calculates front pulse neuron by attenuation function and inputs information, can support with time depth
Space-time Pulse neural network model can be greatly improved compared to the nerual network technique scheme that time depth is only one
The space time information code capacity of impulsive neural networks enriches the application space of impulsive neural networks.
In one of the embodiments, by reading random threshold value mask current potential and threshold bias, and receive configuration deposit
The Configuration Values that device provides determine the threshold potential, so that neuron provides pulse tip formation and has the random of certain probability
Property, no matter film potential is either with or without being more than that fixed threshold biases, due to depositing there are one the threshold value random superposition amount that can just bearing
It is likely to provide pulse in, the pericaryon, improves the computing capability and information processing energy of impulsive neural networks model
Power
Enabled mark is provided by setting in one of the embodiments, and provides trigger flag, determines current PRF mind
Through first output information, so that the controllability of the output of spiking neuron is higher, the neuron that enabler flags can be configured with is provided
Do not allow to provide data, and be only used as intermediate auxiliary and calculate neuron, this is for some function for needing multi-neuron cooperation to complete
It can be very important.
The present invention also provides a kind of adaptive threshold neuronal messages processing systems, comprising:
Front pulse neuron output information receiving module is used for receiving front-end spiking neuron output information;
Current PRF neuronal messages read module, for reading current PRF neuronal messages;
Current PRF neuron output information computing module, for according to the front pulse neuron output information and institute
Current PRF neuronal messages are stated, current PRF neuron output information is calculated;
Current adaptive threshold computing module, for reading current adaptive threshold variable and threshold potential, and according to institute
Current adaptive threshold variable and the threshold potential are stated, current adaptive threshold is calculated;
Current PRF neuron output information output module, for whether judging the current PRF neuron output information
More than or equal to the adaptive threshold, if so, exporting the current PRF neuron output information, and adaptive according to first
Threshold value is answered to update current adaptive threshold variable described in model modification, if it is not,
It does not export the current PRF neuron output information then, and model modification institute is updated according to the second adaptive threshold
State current adaptive threshold variable.
It does not export the current PRF neuron output information then, and model modification institute is updated according to the second adaptive threshold
Stating current adaptive threshold variable, the front pulse neuron output information includes: front end arteries and veins in one of the embodiments,
Rush the pulse tip formation of neuron output, the connection weight index of front pulse neuron and current PRF neuron;
The current PRF neuronal messages include: current time window width, pulse tip formation sequence in current time window
Column, history film potential information and film potential leakage information;
The current PRF neuron output information computing module, comprising:
Spiking neuron connection weight reading unit, for according to the front pulse neuron and current PRF neuron
Connection weight index, read front pulse neuron and current PRF neuron connection weight;
Pulse tip formation sequence updating unit in time window, the pulse for being exported according to the front pulse neuron
Pulse tip formation sequence in tip formation and the current time window, updates pulse tip formation in the current time window
Sequence obtains pulse tip formation renewal sequence in current time window;
Front pulse neuron inputs information calculating unit, for according to the current time window width, the front end arteries and veins
The connection weight for rushing neuron Yu current PRF neuron calculates front pulse neuron by attenuation function and inputs information;
Spiking neuron output information computing unit, for inputting information according to the front pulse neuron, described working as
Pulse tip formation renewal sequence, the history film potential information, the film potential reveal information in preceding time window, pass through pulse
Neural relationship calculates current PRF neuron output information.
In one of the embodiments, further include:
Triggering determination unit is provided, for determining that providing trigger flag information is to provide triggering, the granting trigger flag
Information includes providing triggering or providing not trigger;
Trigger action unit is provided, refractory period timer is used for, and it is preset multiple for updating the history film potential information
Position film potential information.
In one of the embodiments, further include:
Determination unit is triggered when providing, the granting trigger flag information determined is to provide not trigger;
Not trigger action unit is provided, the current time for reading refractory period width and refractory period timer walks;According to
The current time of the refractory period width and refractory period timer step, judges current time whether in refractory period, if working as
The preceding time in the refractory period, by cumulative one time step of timing of the refractory period timer, does not update the history film electricity
Position information;If current time not in the refractory period, by cumulative one time step of timing of the refractory period timer, and updates
The history film potential information is the current PRF neuron output information.
The current adaptive threshold computing module in one of the embodiments, comprising:
Threshold information receiving unit, for reading random threshold value mask current potential, threshold bias and random threshold value;
Threshold value random superposition amount acquiring unit, for by the random threshold value and the random threshold value mask current potential carry out by
Position and operation, obtain threshold value random superposition amount;
Threshold potential determination unit, for determining the threshold according to the threshold value random superposition amount and the threshold bias
It is worth current potential;
Current adaptive threshold determination unit is used for according to the threshold potential and the current adaptive threshold variable,
Determine the current adaptive threshold.
The current PRF neuron output information output module in one of the embodiments, comprising:
First adaptive threshold updating unit provides threshold delta and the current adaptive threshold variable for reading;
According to preset attenuation constant and the current adaptive threshold variable, first threshold is calculated;The granting threshold delta is folded
The first threshold is added to, second threshold is obtained;The current adaptive threshold variable is updated according to the second threshold.
The current PRF neuron output information output module in one of the embodiments, comprising:
Second adaptive threshold updating unit, for reading the current adaptive threshold variable;According to described preset
Attenuation constant and the current adaptive threshold variable calculate third threshold value, according to the third threshold value update it is described it is current from
Adapt to thresholding variables.
The current PRF neuron output information output module in one of the embodiments, comprising:
Enabled mark reading unit is provided, is used to read and provides enabled mark, the enabled mark of granting includes allowing to send out
It puts data or does not allow to provide data, when the granting, which enables to be identified as, to be allowed to provide data,
Trigger flag Information reading unit is provided, for reading the granting trigger flag information, when the granting triggers
Flag information is when providing triggering;
Current PRF neuron output information output unit, for exporting the current PRF neuron output information.
Adaptive threshold neuronal messages processing system provided by the present invention, in current PRF light nerve output information
In calculating, by reading current adaptive threshold variable and threshold potential, current adaptive threshold is calculated;According to current adaptive
Answer threshold value, it is determined whether output current PRF neuron output information, and determine the model for updating current adaptive dependent variable.This hair
The neuronal messages processing system of adaptive threshold provided by bright can make the neuron threshold value currently provided increase, under
Secondary granting difficulty increases;And the neuron threshold value that do not provide currently reduces, next time provides difficulty and reduces.In this way, can effectively
The granting frequency of each neuron in the whole network that weighs, so that each neuron can play a role when handling information, greatly
The big information processing capability for improving impulsive neural networks.
According to the pulse tip formation of front pulse neuron output and described work as in one of the embodiments,
Pulse tip formation sequence in preceding time window updates pulse tip formation sequence in the current time window, obtains current time
Pulse tip formation renewal sequence in window, according to the current time window width, the front pulse neuron and current PRF
The connection weight of neuron calculates front pulse neuron by attenuation function and inputs information, can support with time depth
Space-time Pulse neural network model can be greatly improved compared to the nerual network technique scheme that time depth is only one
The space time information code capacity of impulsive neural networks enriches the application space of impulsive neural networks.
In one of the embodiments, by reading random threshold value mask current potential and threshold bias, and receive configuration deposit
The Configuration Values that device provides determine the threshold potential, so that neuron provides pulse tip formation and has the random of certain probability
Property, no matter film potential is either with or without being more than that fixed threshold biases, due to depositing there are one the threshold value random superposition amount that can just bearing
It is likely to provide pulse in, the pericaryon, improves the computing capability and information processing energy of impulsive neural networks model
Power.
Enabled mark is provided by setting in one of the embodiments, and provides trigger flag, determines current PRF mind
Through first output information, so that the controllability of the output of spiking neuron is higher, the neuron that enabler flags can be configured with is provided
Do not allow to provide data, and be only used as intermediate auxiliary and calculate neuron, this is for some function for needing multi-neuron cooperation to complete
It can be very important.
Detailed description of the invention
Fig. 1 is the flow diagram of the adaptive threshold neural network information processing method of one embodiment;
Fig. 2 is the flow diagram of the adaptive threshold neural network information processing method of another embodiment;
Fig. 3 is the flow diagram of the adaptive threshold neural network information processing method of another embodiment;
Fig. 4 is the flow diagram of the adaptive threshold neural network information processing method of further embodiment;
Fig. 5 is the flow diagram of the adaptive threshold neural network information processing method of further embodiment;
Fig. 6 is the flow diagram of the adaptive threshold neural network information processing method of further embodiment;
Fig. 7 is the structural schematic diagram of the adaptive threshold neural network information processing system of one embodiment;
Fig. 8 is the structural schematic diagram of the adaptive threshold neural network information processing system of another embodiment;
Fig. 9 is the structural schematic diagram of the adaptive threshold neural network information processing system of another embodiment;
Figure 10 is the structural schematic diagram of the adaptive threshold neural network information processing method, system of further embodiment.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, right with reference to the accompanying drawings and embodiments
The present invention is further elaborated.It should be appreciated that described herein, specific examples are only used to explain the present invention, not
For limiting the present invention.
Fig. 1 is the flow diagram of the adaptive threshold neural network information processing method of one embodiment, as shown in Figure 1
Adaptive threshold neural network information processing method, comprising:
Step S100, receiving front-end spiking neuron output information.
Specifically, the front pulse neuron output information, is the arteries and veins for the front end connecting with current PRF neuron
Rush the pulse information of neuron output.
Step S200 reads current PRF neuronal messages.
Specifically, the current PRF neuronal messages, the front pulse neuron including the storage of Current neural member is sent
History pulse information sequence etc..
Step S300 is calculated according to the front pulse neuron output information and the current PRF neuronal messages
Current PRF neuron output information.
Specifically, the pulse information that current PRF neuron is exported according to the front pulse neuron received, and read
The current PRF neuronal messages arrived calculate current PRF neuron output information.
Step S400 reads current adaptive threshold variable and threshold potential, and is become according to the current adaptive threshold
Amount and the threshold potential, calculate current adaptive threshold.
Specifically, the adaptive threshold variable, is calculated according to a time step on current PRF neuron
Current PRF neuron output information (i.e. spike), if finally sent determination.If the pulse information of a upper time step
It is not sent, then the adaptive threshold variable continues to decay on the basis of a upper time step, so that next time step
Pulse train granting difficulty reduce, if the pulse information of a upper time step is sent, the adaptive threshold variable
After decaying on the basis of a upper time step, increase a fixed increment, so that the pulse train of next time step
Difficulty is provided to increase.
Step S500, when whether the current PRF neuron output information is greater than or equal to the adaptive threshold,
If so, meeting step S600, step S700 is otherwise skipped to.
Specifically, judging whether current PRF neuron output information is greater than or equal to the adaptive threshold, with after an action of the bowels
Continuous step determines whether to send the current PRF neuron output information according to judging result.
Step S600 exports the current PRF neuron output information, and according to the first adaptive threshold more new model
Update the current adaptive threshold variable.
Specifically, working as described in transmission if current PRF neuron output information is greater than or equal to the adaptive threshold
Prepulse neuron output information becomes according to current adaptive threshold described in first adaptive threshold update model modification
Amount, so that the current adaptive threshold variable in the calculating of next time step, increases a fixed increment, under increase
The granting difficulty of the calculated current PRF neuron output information of one time step.
Step S700 does not export the current PRF neuron output information then, and is updated according to the second adaptive threshold
Current adaptive threshold variable described in model modification.
Specifically, not sending the current arteries and veins if current PRF neuron output information is less than the adaptive threshold
Rush neuron output information, according to second adaptive threshold update model modification described in current adaptive threshold variable, with
Make the current adaptive threshold variable in the calculating of next time step, continue to decay, reduces next time step and calculate
The granting difficulty of current PRF neuron output information out.
Adaptive threshold neuronal messages processing method provided by the present invention, in current PRF neuron output information
In calculating, by reading current adaptive threshold variable and threshold potential, current adaptive threshold is calculated;According to current adaptive
Answer threshold value, it is determined whether output current PRF neuron output information, and determine the model for updating current adaptive dependent variable.This hair
The spiking neuron information processing method of adaptive threshold provided by bright, can make the nerve for currently having provided output information
First threshold value increases, and provides difficulty next time and increases;And the neuron threshold value for not providing output information currently reduces, next time provides difficulty
It reduces, the granting frequency of each neuron in efficient balance whole network, so that each neuron can be when handling information
It plays a role, greatly improves the information processing capability of impulsive neural networks.
Fig. 2 is the flow diagram of the adaptive threshold neural network information processing method of another embodiment, such as Fig. 2 institute
The adaptive threshold neural network information processing method shown, comprising:
Step S100b, pulse tip formation, front pulse neuron and the current arteries and veins of the output of receiving front-end spiking neuron
Rush the connection weight index of neuron.
Specifically, the connection weight of the front pulse neuron and current PRF neuron indexes, it is front end neuron
The weight index sent together with the front pulse neuron output information, is used to indicate the extraction of Current neural member weight.
The pulse tip formation of the front pulse neuron output, the pulse tip signal sent for front pulse neuron
(spike)。
Step S200b reads current time window width, pulse tip formation sequence, history film potential in current time window
Information and film potential leakage information.
Specifically, pulse tip formation sequence in the current time window, refers in the current time window width, it will
The pulse tip formation that past a certain range of time step receives, the information sequence successively cached in chronological order.
Step S300b is indexed according to the connection weight of the front pulse neuron and current PRF neuron, before reading
Hold the connection weight of spiking neuron and current PRF neuron.
Specifically, the connection weight of the front pulse neuron and current PRF neuron indexes, it is an address letter
Breath, Current neural member are indexed according to the connection weight of the front pulse neuron and current PRF neuron that receive,
In memory in Current neural member, the connection weight of front pulse neuron Yu current PRF neuron is read, according to institute
The connection weight information stated, can be by the output information of front end neuron, in the calculating for participating in Current neural member output information
Cheng Zhong more accurately reflects the weight of the output information of front end neuron, carries richer information.
Step S400b, according in the pulse tip formation and the current time window of front pulse neuron output
Pulse tip formation sequence updates pulse tip formation sequence in the current time window, obtains pulse point in current time window
Client information renewal sequence.
Specifically, the pulse tip formation sequence, walks in the operation of each spiking neuron, one is stored in sequence head
After new pulse tip formation, the pulse tip formation on a sequence tail position is deleted, updates primary pulse tip sequence
Column.
Step S500b updates sequence according to pulse tip formation in the current time window width, the current time window
Column calculate front pulse neuron by attenuation function and input information.
Specifically, utilizingCalculate the front pulse neuron input information, wherein Tw
For the time window width, δjAfter providing spike in current time window for front end neuron j, in the current time window
Time step in pulse tip formation renewal sequence.(Δ t) is an attenuation function to K, is reduced rapidly as Δ t increases.
Step S600b inputs information, the front pulse neuron and current PRF according to the front pulse neuron
The connection weight of neuron, the history film potential information, the film potential reveal information, calculate mould by spiking neuron
Type calculates current PRF neuron output information.
Specifically, indicating the calculating of front pulse neuron input information using following formula:
Wherein WijFor the connection weight of the front pulse neuron j and current PRF neuron i, TwFor the time window
Width, δjAfter providing spike in current time window for front end neuron j, the pulse tip formation in the current time window
Time step in renewal sequence.(Δ t) is an attenuation function to K, is reduced rapidly as Δ t increases.VleakFor leakage value letter
Breath.Basic model at cell space can simplify are as follows:
VSNN=f (V+Vinput+Vleak)
It provides model and reset model is constant, wherein V is the history film potential information that memory saves,
VinputIt is the current input clapped and added up, is equivalent to above-mentioned VleakTo leak value information.
In the present embodiment, according to the pulse tip formation of front pulse neuron output and the current time
Pulse tip formation sequence in window updates pulse tip formation sequence in the current time window, obtains arteries and veins in current time window
Tip formation renewal sequence is rushed, according to the current time window width, the front pulse neuron and current PRF neuron
Connection weight, by attenuation function calculate front pulse neuron input information, can support the space-time with time depth
Impulsive neural networks model can greatly improve pulse mind compared to the nerual network technique scheme that time depth is only one
Space time information code capacity through network enriches the application space of impulsive neural networks.
Fig. 3 is the flow diagram of the adaptive threshold neural network information processing method of another embodiment, such as Fig. 3 institute
The adaptive threshold neural network information processing method shown, comprising:
Step S100c calculates current PRF neuron output information and adaptive threshold.
Step S200c, judges whether the current PRF neuron output information is greater than or equal to the adaptive threshold
When, determine that providing trigger flag information is to provide triggering according to the comparison result, the granting trigger flag information includes hair
It puts triggering or provides and do not trigger, when determining that providing trigger flag information is to provide triggering, meet step S300c, provided when determining
Trigger flag information is when providing not trigger, to skip to step S400c.
Specifically, be compared according to the adaptive threshold current potential with the current PRF neuron output information, and
It is determined according to comparison result and provides trigger flag information.The only described current PRF neuron output information is greater than described adaptive
When threshold potential, the current PRF neuron output information can just be sent.
Step S300c resets refractory period timer, and updating the history film potential information is preset reset film potential
Information.
Specifically, when the granting trigger flag information is to provide triggering, the current PRF neuron output information
It is sent, after refractory period timer is reset, recalculates refractory period, and updating the history film potential information is preset film
Electrical potential information, and the history film potential information update, according to the reset types of configuration, film potential is reset to work as by selectivity
Preceding film potential, current film potential and threshold potential difference or fixed reset voltage.
Step S400c reads the current time step of refractory period width and refractory period timer.
Specifically, when the granting trigger flag information is to provide not trigger, the current PRF neuron output letter
Breath is not sent, and further whether judgement is current in refractory period.The refractory period width is the duration range of refractory period, described
The timing in the way of time step of refractory period timer.
Step S500c is walked according to the current time of the refractory period width and the refractory period timer, when judging current
Between whether in refractory period, if current time in the refractory period, meets step S600c, otherwise skip to step S700c.
Specifically, the cumulative calculation walked according to the current time of the refractory period timer, it can be determined that go out current time
Whether step is also in refractory period.
Cumulative one time step of timing of the refractory period timer is not updated the history film potential letter by step S600c
Breath.
Specifically, if current time in the refractory period, according to the bionical feature of impulsive neural networks, not to the arteries and veins
It rushes neural output information and carries out any response, not more new historical film potential information, the history film potential information, when being next
The spiking neuron of spacer step needs the information read, i.e., in refractory period, this calculated spiking neuron output information is not
Participate in the calculating of next time step.
Step S700c by cumulative one time step of timing of the refractory period timer, and updates the history film potential letter
Breath is the current PRF neuron output information.
Specifically, be then the current PRF neuron output information by the history film potential information such as outside not answering,
Participate in the calculating of next time step.
In the present embodiment, by adaptive threshold current potential, when so that neuron providing pulse tip formation with upper one
Whether spacer step has provided the current PRF neuron output information correlation, can be with each neuron in efficient balance whole network
Granting frequency greatly improve the letter of impulsive neural networks so that each neuron can play a role when handling information
Cease processing capacity.
The output current PRF neuron output information in one of the embodiments, including reading granting make
It can identify, the enabled mark of granting includes allowing granting data or not allowing granting data, provides enabled be identified as when described
When allowing to provide data, the granting trigger flag information is read, when the granting trigger flag information is to provide triggering;It is defeated
The current PRF neuron output information out.
In the present embodiment, enabled mark is provided by setting and provides trigger flag, determine that current PRF neuron is defeated
Information out, so that the controllability of the output of spiking neuron is higher, the neuron that providing enabler flags can be configured with does not allow
Data are provided, and is only used as intermediate auxiliary and calculates neuron, this is for some function right and wrong for needing multi-neuron cooperation to complete
It is often necessary.
Fig. 4 is the flow diagram of the adaptive threshold neural network information processing method of further embodiment, such as Fig. 4 institute
The adaptive threshold neural network information processing method shown, comprising:
Step S410 reads random threshold value mask current potential, threshold bias, current adaptive threshold variable and random threshold value.
Specifically, the current adaptive threshold variable, be when a task starts, it is preset adaptive according to one
Threshold value initial value, preset attenuation coefficient and provide threshold delta, gradually decay or increases according to each time step, obtain oneself
Adapt to thresholding variables value.The current adaptive threshold variable, the current time step read for current PRF neuron from
Adapt to thresholding variables.The random threshold value provides a random number seed by configuration register, passes through pseudorandom number generator
It generates.
The random threshold value and the random threshold value mask current potential are carried out step-by-step and operation, obtain threshold value by step S420
Random superposition amount.
Specifically, the random threshold value mask current potential is used for the range of threshold limit increment.
Step S430 determines the threshold potential according to the threshold value random superposition amount and the threshold bias.
Specifically, the threshold value random superposition amount is added with the threshold bias, the threshold potential is obtained.
Step S440 is determined described current adaptive according to the threshold potential and the current adaptive threshold variable
Threshold value.
Specifically, the threshold potential and the current adaptive threshold addition of variables are determined described current adaptive
Threshold value.Wherein, by the threshold value random superposition amount and preset threshold bias Vth0It is added, generates real threshold potential Vth.Its
In, the seed of pseudorandom number generator is by configuration register VseedIt provides.Mask current potential VmaskModel for threshold limit increment
It encloses: if Vmask=0, then threshold value random superposition amount is also 0, and it is fixed threshold granting, fixed threshold V that Firing Patterns, which are degenerated,th0;
If Vmask≠ 0, then Firing Patterns are the granting of part probability threshold value.As extreme case Vth0=0, then Firing Patterns are full-probability
Threshold value granting.
In the present embodiment, it by reading random threshold value mask current potential and threshold bias, and receives configuration register and provides
Configuration Values, determine the threshold potential so that neuron provide pulse tip formation have certain probability randomness, no matter
Film potential is either with or without being more than that fixed threshold biases, since there are one the presence for the threshold value random superposition amount that can just bearing, the minds
It is likely to provide pulse through first cell space, improves the computing capability and information processing capability of impulsive neural networks model.
Fig. 5 is the flow diagram of the adaptive threshold neural network information processing method of further embodiment, such as Fig. 5 institute
The adaptive threshold neural network information processing method shown is in Fig. 1, according to the first adaptive threshold update model modification
The detailed step of current adaptive threshold variable part, comprising:
Step S610 reads and provides threshold delta and the current adaptive threshold variable.
Specifically, the granting threshold delta is preset value, can be set according to the demand of task.
Step S620 calculates first threshold according to preset attenuation constant and the current adaptive threshold variable.
Specifically, the current adaptive threshold variable that will be read, calculates according to preset attenuation coefficient, the first threshold is obtained
Value, the first threshold are that this time step executes the adaptive threshold variable after decaying.
The granting threshold delta is superimposed to the first threshold, obtains second threshold by step S630.
Specifically, after this time step executes decaying, then will be first described in the preset granting threshold delta superposition value
Threshold value obtains second threshold.
Step S640 updates the current adaptive threshold variable according to the second threshold.
Specifically, being the second threshold by the current adaptive threshold variable update.Because of the second threshold, be
After adaptive dependent variable performs decay calculation, and it is superimposed preset granting threshold delta, so that next time step was read
Adaptive threshold increases, and improves the granting difficulty of the current PRF neuron output information of next time step.
First adaptive threshold more new model provided in the present embodiment is sent in current PRF neuron output information
In the case where, after current adaptive threshold is executed decaying, it is superimposed a preset increment, so that next time step is current
The granting difficulty of spiking neuron output information increases.Flexible characteristic is added for threshold value in this way, can be made entire
The granting rate of network is relatively uniform, and each neuron has an opportunity to learn respective input receptive field out.
Fig. 6 is the flow diagram of the adaptive threshold neural network information processing method of further embodiment, such as Fig. 6 institute
The adaptive threshold neural network information processing method shown is in Fig. 1, according to the second adaptive threshold update model modification
The detailed step of current adaptive threshold variable part, comprising:
Step S710 reads the current adaptive threshold variable.
Specifically, not needing then to improve adaptive threshold since current PRF neuron output information is not sent.
Step S720 calculates third threshold value according to the preset attenuation constant and the current adaptive threshold variable.
Specifically, the current adaptive threshold variable that will be read, calculates according to preset attenuation coefficient, third threshold is obtained
Value, the third threshold value are that this time step executes the adaptive threshold variable after decaying.
Step S730 updates the current adaptive threshold variable according to the third threshold value.
Specifically, that is, third threshold value is set as current adaptive threshold increment by the adaptive dependent variable after decaying, reduce
The granting difficulty of the current PRF neuron output information of next time step.
Second adaptive threshold more new model provided in the present embodiment is not sent out in current PRF neuron output information
In the case where sending, current adaptive threshold is executed into decaying, so that the current PRF neuron output information of next time step
Granting difficulty reduce.Flexible characteristic is added for threshold value in this way, the granting rate of whole network can be made more equal
Even, each neuron has an opportunity to learn respective input receptive field out.
Fig. 7 is the structural schematic diagram of the adaptive threshold neural network information processing system of one embodiment, as shown in Figure 7
Adaptive threshold neural network information processing system, comprising:
Front pulse neuron output information receiving module 100 is used for receiving front-end spiking neuron output information;It is described
Front pulse neuron output information include: front pulse neuron output pulse tip formation, front pulse neuron with
The connection weight of current PRF neuron indexes.
Current PRF neuronal messages read module 200, for reading current PRF neuronal messages;The current PRF
Neuronal messages include: current time window width, pulse tip formation sequence, history film potential information and film in current time window
Potential leakage information.
Current PRF neuron output information computing module 300, for according to the front pulse neuron output information
With the current PRF neuronal messages, current PRF neuron output information is calculated;
Current adaptive threshold computing module 400, for reading current adaptive threshold variable and threshold potential, and according to
The current adaptive threshold variable and the threshold potential, calculate current adaptive threshold;
Current PRF neuron output information output module 500 is big for working as the current PRF neuron output information
When the adaptive threshold, the current PRF neuron output information is exported, and according to the first adaptive threshold
Current adaptive threshold variable described in model modification is updated, does not otherwise export the current PRF neuron output information then, and
According to current adaptive threshold variable described in the second adaptive threshold update model modification.
Adaptive threshold neuronal messages processing system provided by the present invention, in current PRF light nerve output information
In calculating, by reading current adaptive threshold variable and threshold potential, current adaptive threshold is calculated;According to current adaptive
Answer threshold value, it is determined whether output current PRF neuron output information, and determine the model for updating current adaptive dependent variable.This hair
The neuronal messages processing system of adaptive threshold provided by bright can make the neuron threshold value currently provided increase, under
Secondary granting difficulty increases;And the neuron threshold value that do not provide currently reduces, next time provides difficulty and reduces.In this way, can effectively
The granting frequency of each neuron in the whole network that weighs, so that each neuron can play a role when handling information, greatly
The big information processing capability for improving impulsive neural networks.
Fig. 8 is the structural schematic diagram of the adaptive threshold neural network information processing system of another embodiment, such as Fig. 8 institute
The adaptive threshold neural network information processing system shown is the output information computing module of current PRF neuron described in Fig. 7
300, comprising:
Spiking neuron connection weight reading unit 100b, for according to the front pulse neuron and current PRF mind
Connection weight through member indexes, and reads the connection weight of front pulse neuron and current PRF neuron;
Pulse tip formation sequence updating unit 200b in time window, for what is exported according to the front pulse neuron
Pulse tip formation sequence in pulse tip formation and the current time window, updates pulse tip in the current time window
Information sequence obtains pulse tip formation renewal sequence in current time window;
Front pulse neuron input information calculating unit 300b, for according to the current time window width, it is described before
The connection weight for holding spiking neuron and current PRF neuron calculates front pulse neuron input letter by attenuation function
Breath;
Spiking neuron output information computing unit 400b, for inputting information, institute according to the front pulse neuron
Pulse tip formation renewal sequence in current time window, the history film potential information, film potential leakage information are stated, is passed through
Spiking neuron computation model calculates current PRF neuron output information.
In the present embodiment, according to the pulse tip formation of front pulse neuron output and the current time
Pulse tip formation sequence in window updates pulse tip formation sequence in the current time window, obtains arteries and veins in current time window
Tip formation renewal sequence is rushed, according to the current time window width, the front pulse neuron and current PRF neuron
Connection weight, by attenuation function calculate front pulse neuron input information, can support the space-time with time depth
Impulsive neural networks model can greatly improve pulse mind compared to the nerual network technique scheme that time depth is only one
Space time information code capacity through network enriches the application space of impulsive neural networks.
Fig. 9 is the structural schematic diagram of the adaptive threshold neural network information processing system of another embodiment, such as Fig. 9 institute
The adaptive threshold neural network information processing system shown is current PRF neuron output information output module described in Fig. 7,
Include:
Triggering determination unit 100c is provided, for determining that providing trigger flag information is to provide triggering, the granting triggering
Flag information includes providing triggering or providing not trigger;
Trigger action unit 200c is provided, is used for refractory period timer, and it is default for updating the history film potential information
Reset film potential information.
Determination unit 100c is triggered when providing, the granting trigger flag information determined is to provide not trigger;
Not trigger action unit 300c is provided, the current time for reading refractory period width and refractory period timer walks;
According to the current time of the refractory period width and the refractory period timer walk, judge current time whether in refractory period,
If current time in the refractory period, by cumulative one time step of timing of the refractory period timer, does not update the history
Film potential information;If current time is not in the refractory period, by cumulative one time step of timing of the refractory period timer, and
Updating the history film potential information is the current PRF neuron output information.
First adaptive threshold updating unit 400c provides threshold delta and the current adaptive threshold change for reading
Amount;According to preset attenuation constant and the current adaptive threshold variable, first threshold is calculated;By the granting threshold delta
It is superimposed to the first threshold, obtains second threshold;The current adaptive threshold variable is updated according to the second threshold.
Second adaptive threshold updating unit 500c, for reading the current adaptive threshold variable;According to described pre-
If attenuation constant and the current adaptive threshold variable, calculate third threshold value, according to the third threshold value update described in work as
Preceding adaptive threshold variable.
Enabled mark reading unit 600c is provided, is used to read and provides enabled mark, the enabled mark of granting includes permitting
Perhaps data are provided or do not allow to provide data, when the granting, which enables to be identified as, to be allowed to provide data,
Trigger flag Information reading unit 700c is provided, for reading the granting trigger flag information, when the granting
Trigger flag information is when providing triggering.
Current PRF neuron output information output unit 800c, for exporting the current PRF neuron output letter
Breath.
The more new model of adaptive threshold provided in the present embodiment is sent in current PRF neuron output information
In the case of, after current adaptive threshold is executed decaying, it is superimposed a preset increment, so that the current arteries and veins of next time step
The granting difficulty for rushing neuron output information increases.It, will be current in the case where current PRF neuron output information is not sent
Adaptive threshold executes decaying, so that the granting difficulty of the current PRF neuron output information of next time step reduces.It is logical
Crossing this mode is that threshold value adds flexible characteristic, the granting rate of whole network can be made relatively uniform, each neuron has
Chance learns respective input receptive field out.
Figure 10 is the structural schematic diagram of the adaptive threshold neural network information processing method, system of further embodiment, is such as schemed
Adaptive threshold neural network information processing system shown in 10 is the current adaptive threshold computing module 400 in Fig. 7, packet
It includes:
Threshold information receiving unit 410, for reading random threshold value mask current potential, threshold bias and random threshold value;
Threshold value random superposition amount acquiring unit 420, for by the random threshold value and the random threshold value mask current potential into
Row step-by-step and operation, obtain threshold value random superposition amount;
Threshold potential determination unit 430, described in determining according to the threshold value random superposition amount and the threshold bias
Threshold potential;
Current adaptive threshold determination unit 440, for being become according to the threshold potential and the current adaptive threshold
Amount, determines the current adaptive threshold.
In the present embodiment, it by reading random threshold value mask current potential and threshold bias, and receives configuration register and provides
Configuration Values, determine the threshold potential so that neuron provide pulse tip formation have certain probability randomness, no matter
Film potential is either with or without being more than that fixed threshold biases, since there are one the presence for the threshold value random superposition amount that can just bearing, the minds
It is likely to provide pulse through first cell space, improves the computing capability and information processing capability of impulsive neural networks model.
Each technical characteristic of embodiment described above can be combined arbitrarily, for simplicity of description, not to above-mentioned reality
It applies all possible combination of each technical characteristic in example to be all described, as long as however, the combination of these technical characteristics is not deposited
In contradiction, all should be considered as described in this specification.
The embodiments described above only express several embodiments of the present invention, and the description thereof is more specific and detailed, but simultaneously
It cannot therefore be construed as limiting the scope of the patent.It should be pointed out that coming for those of ordinary skill in the art
It says, without departing from the inventive concept of the premise, various modifications and improvements can be made, these belong to protection of the invention
Range.Therefore, the scope of protection of the patent of the invention shall be subject to the appended claims.
Claims (16)
1. a kind of adaptive threshold neuronal messages processing method, which is characterized in that the described method includes:
Receiving front-end spiking neuron output information;
Read current PRF neuronal messages;
According to the front pulse neuron output information and the current PRF neuronal messages, current PRF neuron is calculated
Output information;
Current adaptive threshold variable and threshold potential are read, and according to the current adaptive threshold variable and threshold value electricity
Position, calculates current adaptive threshold;
Judge whether the current PRF neuron output information is greater than or equal to the adaptive threshold, if so, output institute
Current PRF neuron output information is stated, and the current adaptive threshold according to the first adaptive threshold update model modification becomes
Amount, if it is not,
The current PRF neuron output information is not exported, and current according to the second adaptive threshold update model modification
Adaptive threshold variable.
2. adaptive threshold neuronal messages processing method according to claim 1, it is characterised in that:
The front pulse neuron output information includes: the pulse tip formation of front pulse neuron output, front pulse
The connection weight of neuron and current PRF neuron indexes;
The current PRF neuronal messages include: current time window width, pulse tip formation sequence in current time window, go through
History film potential information and film potential leakage information;
Then according to the front pulse neuron output information and the current PRF neuronal messages, current PRF nerve is calculated
First output information, comprising:
Indexed according to the connection weight of the front pulse neuron and current PRF neuron, read front pulse neuron with
The connection weight of current PRF neuron;
According to pulse tip formation sequence in the pulse tip formation and the current time window of front pulse neuron output
Column update pulse tip formation sequence in the current time window, obtain pulse tip formation renewal sequence in current time window;
According to pulse tip formation renewal sequence in the current time window width, the current time window, pass through attenuation function
It calculates front pulse neuron and inputs information;
Information, the connection weight of the front pulse neuron and current PRF neuron are inputted according to the front pulse neuron
Weight, the history film potential information, the film potential reveal information, by spiking neuron computation model, calculate current PRF
Neuron output information.
3. adaptive threshold neuronal messages processing method according to claim 2, which is characterized in that when the current arteries and veins
When rushing neuron output information more than or equal to the adaptive threshold, the output current PRF neuron output letter
Breath, and the current adaptive threshold variable according to the first adaptive threshold update model modification, further includes:
Determine that providing trigger flag information is to provide triggering, the granting trigger flag information includes providing triggering or providing not touch
Hair;
Refractory period timer is resetted, and updating the history film potential information is preset reset film potential information.
4. adaptive threshold neuronal messages processing method according to claim 3, which is characterized in that described not export institute
Current PRF neuron output information is stated, and the current adaptive threshold according to the second adaptive threshold update model modification becomes
Amount, further includes:
Determine that the granting trigger flag information does not trigger to provide;
Read the current time step of refractory period width and refractory period timer;
It is walked according to the current time of the refractory period width and the refractory period timer, judges current time whether in refractory period
It is interior, if current time in the refractory period, by cumulative one time step of timing of the refractory period timer, does not update described go through
History film potential information;
If current time not in the refractory period, by cumulative one time step of timing of the refractory period timer, and updates institute
Stating history film potential information is the current PRF neuron output information.
5. adaptive threshold neuronal messages processing method according to claim 1, which is characterized in that described to read currently
Adaptive threshold variable and threshold potential, and according to the current adaptive threshold variable and the threshold potential, it calculates current
Adaptive threshold, comprising:
Read random threshold value mask current potential, threshold bias, current adaptive threshold variable and random threshold value;
The random threshold value and the random threshold value mask current potential are subjected to step-by-step and operation, obtain threshold value random superposition amount;
According to the threshold value random superposition amount and the threshold bias, the threshold potential is determined;
According to the threshold potential and the current adaptive threshold variable, the current adaptive threshold is determined.
6. adaptive threshold neuronal messages processing method according to claim 1, which is characterized in that described according to first
Adaptive threshold updates current adaptive threshold variable described in model modification, comprising:
It reads and provides threshold delta and the current adaptive threshold variable;
According to preset attenuation constant and the current adaptive threshold variable, first threshold is calculated;
The granting threshold delta is superimposed to the first threshold, obtains second threshold;
The current adaptive threshold variable is updated according to the second threshold.
7. adaptive threshold neuronal messages processing method according to claim 6, which is characterized in that described according to second
Adaptive threshold updates current adaptive threshold variable described in model modification, comprising:
Read the current adaptive threshold variable;
According to the preset attenuation constant and the current adaptive threshold variable, third threshold value is calculated,
The current adaptive threshold variable is updated according to the third threshold value.
8. adaptive threshold neuronal messages processing method according to claim 3, which is characterized in that described in the output
Current PRF neuron output information, comprising:
It reads and provides enabled mark, the enabled mark of granting includes allowing to provide data or not allowing granting data, when described
Granting enables to be identified as when allowing to provide data,
The granting trigger flag information is read, when the granting trigger flag information is to provide triggering;
Export the current PRF neuron output information.
9. a kind of adaptive threshold neuronal messages processing system characterized by comprising
Front pulse neuron output information receiving module is used for receiving front-end spiking neuron output information;
Current PRF neuronal messages read module, for reading current PRF neuronal messages;
Current PRF neuron output information computing module, for according to the front pulse neuron output information and described working as
Prepulse neuronal messages calculate current PRF neuron output information;
Current adaptive threshold computing module, works as reading current adaptive threshold variable and threshold potential, and according to described
Preceding adaptive threshold variable and the threshold potential, calculate current adaptive threshold;
Current PRF neuron output information output module, for judging whether the current PRF neuron output information is greater than
Or it is equal to the adaptive threshold, if so, exporting the current PRF neuron output information, and according to the first adaptive thresholding
Value updates current adaptive threshold variable described in model modification, if it is not,
Do not export the current PRF neuron output information then, and according to the second adaptive threshold update model modification described in when
Preceding adaptive threshold variable.
10. adaptive threshold neuronal messages processing system according to claim 9, it is characterised in that:
The front pulse neuron output information includes: the pulse tip formation of front pulse neuron output, front pulse
The connection weight of neuron and current PRF neuron indexes;
The current PRF neuronal messages include: current time window width, pulse tip formation sequence in current time window, go through
History film potential information and film potential leakage information;
The current PRF neuron output information computing module, comprising:
Spiking neuron connection weight reading unit, for the company according to the front pulse neuron and current PRF neuron
Weight index is connect, the connection weight of front pulse neuron and current PRF neuron is read;
Pulse tip formation sequence updating unit in time window, the pulse tip for being exported according to the front pulse neuron
Pulse tip formation sequence in information and the current time window updates pulse tip formation sequence in the current time window,
Obtain pulse tip formation renewal sequence in current time window;
Front pulse neuron inputs information calculating unit, for according to the current time window width, front pulse mind
Connection weight through member with current PRF neuron calculates front pulse neuron by attenuation function and inputs information;
Spiking neuron output information computing unit, for according to the front pulse neuron input information, it is described current when
Between in window pulse tip formation renewal sequence, the history film potential information, the film potential reveal information, pass through pulse nerve
Relationship calculates current PRF neuron output information.
11. adaptive threshold neuronal messages processing system according to claim 10, which is characterized in that further include:
Triggering determination unit is provided, for determining that providing trigger flag information is to provide triggering, the granting trigger flag information
It is not triggered including providing triggering or providing;
Trigger action unit is provided, refractory period timer is used for, and updating the history film potential information is preset reset film
Electrical potential information.
12. adaptive threshold neuronal messages processing system according to claim 11, which is characterized in that further include:
Determination unit is triggered when providing, the granting trigger flag information determined is to provide not trigger;
Not trigger action unit is provided, the current time for reading refractory period width and refractory period timer walks;According to described
The current time of refractory period width and refractory period timer step, judges current time whether in refractory period, if when current
Between in the refractory period, the refractory period timer is added up one time step of timing, does not update history film potential letter
Breath;If current time not in the refractory period, by the refractory period timer add up one time step of timing, and update described in
History film potential information is the current PRF neuron output information.
13. adaptive threshold neuronal messages processing system according to claim 9, which is characterized in that it is described it is current from
Adapt to threshold calculation module, comprising:
Threshold information receiving unit, for reading random threshold value mask current potential, threshold bias and random threshold value;
Threshold value random superposition amount acquiring unit, for will the random threshold value and the random threshold value mask current potential carry out step-by-step and
Operation obtains threshold value random superposition amount;
Threshold potential determination unit, for determining the threshold value electricity according to the threshold value random superposition amount and the threshold bias
Position;
Current adaptive threshold determination unit, for determining according to the threshold potential and the current adaptive threshold variable
The current adaptive threshold.
14. adaptive threshold neuronal messages processing system according to claim 9, which is characterized in that the current arteries and veins
Rush neuron output information output module, comprising:
First adaptive threshold updating unit provides threshold delta and the current adaptive threshold variable for reading;According to
Preset attenuation constant and the current adaptive threshold variable calculate first threshold;The granting threshold delta is superimposed to
The first threshold obtains second threshold;The current adaptive threshold variable is updated according to the second threshold.
15. adaptive threshold neuronal messages processing system according to claim 14, which is characterized in that described to work as
Prepulse neuron output information output module, comprising:
Second adaptive threshold updating unit, for reading the current adaptive threshold variable;According to the preset decaying
Constant and the current adaptive threshold variable calculate third threshold value, are updated according to the third threshold value described current adaptive
Thresholding variables.
16. adaptive threshold neuronal messages processing system according to claim 11, which is characterized in that the current arteries and veins
Rush neuron output information output module, comprising:
Enabled mark reading unit is provided, is used to read and provides enabled mark, the enabled mark of granting includes allowing to provide number
According to or do not allow to provide data, when it is described provide it is enabled be identified as allow to provide data when,
Trigger flag Information reading unit is provided, for reading the granting trigger flag information, when the granting trigger flag
Information is when providing triggering;
Current PRF neuron output information output unit, for exporting the current PRF neuron output information.
Priority Applications (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710042067.5A CN106875005B (en) | 2017-01-20 | 2017-01-20 | Adaptive threshold neuronal messages processing method and system |
US16/478,458 US11551074B2 (en) | 2017-01-20 | 2017-12-05 | Self-adaptive threshold neuron information processing method, self-adaptive leakage value neuron information processing method, system computer device and readable storage medium |
PCT/CN2017/114669 WO2018133570A1 (en) | 2017-01-20 | 2017-12-05 | Self-adaptive threshold neuron information processing method, self-adaptive leakage value neuron information processing method and system, and computer device and readable storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710042067.5A CN106875005B (en) | 2017-01-20 | 2017-01-20 | Adaptive threshold neuronal messages processing method and system |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106875005A CN106875005A (en) | 2017-06-20 |
CN106875005B true CN106875005B (en) | 2019-09-20 |
Family
ID=59158522
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710042067.5A Active CN106875005B (en) | 2017-01-20 | 2017-01-20 | Adaptive threshold neuronal messages processing method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106875005B (en) |
Families Citing this family (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11551074B2 (en) * | 2017-01-20 | 2023-01-10 | Tsinghua University | Self-adaptive threshold neuron information processing method, self-adaptive leakage value neuron information processing method, system computer device and readable storage medium |
CN107609640A (en) * | 2017-10-01 | 2018-01-19 | 胡明建 | A kind of threshold values selects the design method of end graded potential formula artificial neuron |
CN114254106A (en) * | 2020-09-25 | 2022-03-29 | 北京灵汐科技有限公司 | Text classification method, device, equipment and storage medium |
CN112734012B (en) * | 2021-01-07 | 2024-03-05 | 北京灵汐科技有限公司 | Impulse neural network training method, data processing method, electronic equipment and medium |
CN112766469B (en) * | 2021-01-22 | 2024-01-30 | 北京灵汐科技有限公司 | Method and device for generating pulse signal |
CN112949833B (en) * | 2021-03-26 | 2022-09-06 | 北京航空航天大学 | Probability calculation neuron calculation unit and construction method |
CN113449864B (en) * | 2021-07-21 | 2023-08-25 | 北京大学 | Feedback type impulse neural network model training method for image data classification |
CN114861838B (en) * | 2022-07-06 | 2022-09-27 | 武汉理工大学 | Intelligent classification method for pulsatile neural brains based on neuron complex dynamics |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1078053A (en) * | 1992-04-27 | 1993-11-03 | 福建师范大学 | Injection electroluminescent nerve network circuit |
CN102306301A (en) * | 2011-08-26 | 2012-01-04 | 中南民族大学 | Motion identification system by simulating spiking neuron of primary visual cortex |
CN103430186A (en) * | 2012-01-20 | 2013-12-04 | 松下电器产业株式会社 | Learning method for neural network circuit |
CN103778468A (en) * | 2014-01-16 | 2014-05-07 | 北京大学 | RRAM-based new type neural network circuit |
CN105095967A (en) * | 2015-07-16 | 2015-11-25 | 清华大学 | Multi-mode neural morphological network core |
-
2017
- 2017-01-20 CN CN201710042067.5A patent/CN106875005B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1078053A (en) * | 1992-04-27 | 1993-11-03 | 福建师范大学 | Injection electroluminescent nerve network circuit |
CN102306301A (en) * | 2011-08-26 | 2012-01-04 | 中南民族大学 | Motion identification system by simulating spiking neuron of primary visual cortex |
CN103430186A (en) * | 2012-01-20 | 2013-12-04 | 松下电器产业株式会社 | Learning method for neural network circuit |
CN103778468A (en) * | 2014-01-16 | 2014-05-07 | 北京大学 | RRAM-based new type neural network circuit |
CN105095967A (en) * | 2015-07-16 | 2015-11-25 | 清华大学 | Multi-mode neural morphological network core |
Non-Patent Citations (3)
Title |
---|
A highly tunable 65-nm CMOS LIF neuron for a large scale neuromorphic system;Aamir et.al;《IEEE ESSDERC 2016 - 46th European Solid-State Device Research Conference - Lausanne, Switzerland》;20161230(第1期);第1-3页 * |
Bridging the semantic gap: Emulating biological neuronal behaviors with simple digital neurons;Andrew Nere et.al;《 2013 IEEE 19th International Symposium on High Performance Computer Architecture (HPCA)》;20130927(第1期);第1-3页 * |
基于脉冲序列内积的脉冲神经网络监督学习研究;王向文;《中国优秀硕士学位论文全文数据库信息科技辑》;20160715(第07期);第I140-62页 * |
Also Published As
Publication number | Publication date |
---|---|
CN106875005A (en) | 2017-06-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106875005B (en) | Adaptive threshold neuronal messages processing method and system | |
CN106875004B (en) | Composite mode neuronal messages processing method and system | |
CN106875003A (en) | Self adaptation leakage value neuronal messages processing method and system | |
CN110555523B (en) | Short-range tracking method and system based on impulse neural network | |
CN107943874B (en) | Knowledge mapping processing method, device, computer equipment and storage medium | |
CN109919316A (en) | The method, apparatus and equipment and storage medium of acquisition network representation study vector | |
CN110807469B (en) | Knowledge tracking method and system integrating long-time memory and short-time memory with Bayesian network | |
CN107103754A (en) | A kind of road traffic condition Forecasting Methodology and system | |
CN109492765A (en) | A kind of image Increment Learning Algorithm based on migration models | |
JP2016539407A (en) | Causal saliency time inference | |
CN106845633A (en) | Neutral net information conversion method and system | |
CN104424507B (en) | Prediction method and prediction device of echo state network | |
CN104573017B (en) | Identify the method and system of network navy group | |
CN109743196A (en) | It is a kind of based on the network characterisation method across double-layer network random walk | |
CN110472154A (en) | A kind of resource supplying method, apparatus, electronic equipment and readable storage medium storing program for executing | |
CN106875006B (en) | Artificial neuron metamessage is converted to the method and system of spiking neuron information | |
CN113077052A (en) | Reinforced learning method, device, equipment and medium for sparse reward environment | |
CN115293247A (en) | Method for establishing risk identification model, risk identification method and corresponding device | |
US11551074B2 (en) | Self-adaptive threshold neuron information processing method, self-adaptive leakage value neuron information processing method, system computer device and readable storage medium | |
CN105989376B (en) | A kind of hand-written discrimination system neural network based, device and mobile terminal | |
CN108229688A (en) | A kind of information processing method and device based on IRT | |
CN116456480A (en) | Multi-agent collaborative decision-making method based on deep reinforcement learning under communication resource limitation | |
CN116743468A (en) | Dynamic attack path generation method based on reinforcement learning | |
CN116862025A (en) | Model training method, system, client and server node, electronic device and storage medium | |
CN106815638A (en) | The neuronal messages processing method and system that input weight is expanded |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | 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 |