CN110210563A - The study of pattern pulse data space time information and recognition methods based on Spike cube SNN - Google Patents
The study of pattern pulse data space time information and recognition methods based on Spike cube SNN Download PDFInfo
- Publication number
- CN110210563A CN110210563A CN201910481420.9A CN201910481420A CN110210563A CN 110210563 A CN110210563 A CN 110210563A CN 201910481420 A CN201910481420 A CN 201910481420A CN 110210563 A CN110210563 A CN 110210563A
- Authority
- CN
- China
- Prior art keywords
- pulse
- neuron
- pattern
- excitation
- layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses methods and image-recognizing method that a kind of space time information for pattern pulse sequence carries out combination learning, pulse train cell S pike cube and LIF neuron models are established based on impulsive neural networks, and STDP mechanism is used, the synapse weight and excitation threshold connect between each other to each layer neuron of impulsive neural networks learns;Trained model is recycled to carry out image classification identification.The present invention provides new technical solution for the structure design and study, pattern pulse Sequence Learning and identification of impulsive neural networks, while also the pulse data for the bionical vision camera output such as DVS provides new processing method.
Description
Technical field
The invention belongs to the calculating of class brain, impulsive neural networks (SNN), STDP study, image identification technical field, are related to arteries and veins
Rush sequence learning method more particularly to it is a kind of based on spike cube impulsive neural networks to pattern pulse sequence include when
Between and spatial information carry out combination learning method.
Background technique
In recent years, with the fast development of deep learning, extensive artificial neural network is widely used in numerous areas
In the middle.Study and identification to sequence information, can also be realized by depth learning technology.For example, recycling nerve net with numerical value
Network is the time Series Processing method of representative, the sequence information with temporal associativity can be learnt and be handled, extensively
Applying this in natural language processing, machine translation, speech recognition etc. only includes one-dimensional temporal information and one-dimensional language information
In two-dimensional scene, but the 3 D stereo number comprising one-dimensional temporal information and two-dimensional spatial location information this for pulse train
According to Recognition with Recurrent Neural Network does not have more mature study and processing capacity also at present.For another example being with numerical convolution neural network
The image processing method of representative has highly developed information processing capability to flat image, is widely used in image recognition, mesh
The fields such as mark detection and target following, also have the ability of study and identification, but this study and recognition methods to pulse train
It is that temporal accumulation first is carried out to pulse train, is then re-fed into neural network.Although this information processing manner is being known
Relatively high level is not reached in rate, but has substantially been lost the timing information that pulse train carries originally, has not been really to anticipate
The study and recognition methods of pulse train space time information in justice.In addition, also not having biology in study and treatment mechanism
Characteristic, it is also relatively high on calculating energy consumption.
Impulsive neural networks (SNN, Spiking Neuron Networks) are known as " third generation neural network ", compare
In numerical value deep learning neural network popular at present, the more bionical characteristic of impulsive neural networks is embodied in information
In terms for the treatment of mechanism and study mechanism.In terms of information processing mechanism, the information of input and the processing of impulsive neural networks is arteries and veins
Sequence is rushed, it is rather similar in form to corresponding and excitation mechanism of the biological neuron in brain to action potential signal.It is learning
In terms of habit mechanism, the study of impulsive neural networks is based primarily upon STDP mechanism, and STDP mechanism is then considered biological in brain
The mode of learning of neuronal synapse connection.Impulsive neural networks are on information processing mechanism and study mechanism with reference to brain
Mode, therefore, impulsive neural networks are the research branches that class brain calculates.
Although existing most of style pulse neural networks are to the processing of pulse train not on input source to pulse sequence
Column are accumulated, but since its synapse weight is to carry out weight by numerical value neural network to transplant, and numerical value is refreshing mostly
Be through the weight of network learnt by the pulse image after being accumulated to pulse train, so, transplanted by weight
The impulsive neural networks built are only by pulse sequence substantially and without the excavation ability to pulse train timing information
It arranges the spatial information accumulated in time to be identified, it is latent in space time information expression not to play pulse train completely
Power.Pulse nerve there are also a few types is directly learnt to the pulse train that do not accumulate using STDP mechanism, although learning
The object of habit is pulse train itself, but at present such methods learn as a result, synapse weight and mind i.e. between neuron
Static value is still through first excitation threshold.Go processing with the pulse of time response with the static cynapse for not having time response
Sequence is that could not efficiently use the timing information that pulse train is included naturally.
Summary of the invention
In order to overcome the above-mentioned deficiencies of the prior art, the present invention, which provides, is based on spike cube (pulse train unit) pulse
The method of time and spatial information progress combination learning that neural network includes to pattern pulse sequence, are pattern pulse sequence
It practises and identification provides a kind of new technical solution, provide new think of for the structure design and learning method of impulsive neural networks
Road, and further the pulse data for the bionical vision camera output such as DVS provides a kind of new processing method.
Core of the invention is: propose it is a kind of for image procossing based on spike cube and LIF (leaky
Integrate and Fire, leakage current integral excitation) neuron models and use STDP (spiking time dependent
Plasticity, the synaptic plasticity that pulse time relies on) mechanism joins the time of pattern pulse sequence and spatial information
The method for closing study.In the present invention, the minimum time according to time series pattern included in pattern pulse sequence to be identified is long
Degree, the pulse train segment that input picture pulse train is divided at equal intervals and obtains pulse train segment, and will be obtained
It is defined as pulse train unit (Spike cube).Impulsive neural networks, pulse nerve net are built first, in accordance with the framework of attached drawing 2
The neuron of network is indicated using LIF model;It is then ready for pattern pulse sequence to be learned, it is drawn according to certain ratio
It is divided into training set, verifying collection and test set, training set pulse train input pulse neural network is subjected to propagated forward, is successively led to
Cross the first convolutional layer, the second convolutional layer, full articulamentum and classification layer;It is study with spike cube in layer-by-layer communication process
The unit and synapse weight connected between each other using STDP mechanism to neuron and excitation threshold are learnt;It needs especially to say
Bright be convolutional layer synapse weight and excitation threshold has different characteristics at different times form, every study a period of time
Afterwards, the learning process of cynapse is temporarily interrupted, and uses verifying collection pattern pulse sequence inputting to current impulsive neural networks
Carry out image category recognition accuracy test, if reaching expected standard, terminate learning process and save synapse weight and
Excitation threshold continues to learn on this basis if not reaching expected standard, until reaching expected standard or entire instruction
Practice collection pattern pulse sequence to learn to finish;After study, test set pattern pulse is carried out with the impulsive neural networks
Recognition accuracy test.It needs especially to remark additionally, the impulsive neural networks that Fig. 2 is built provide proposed by the present invention
Practise the existing fairly simple impulsive neural networks framework of one kind relied on recognition methods, the pulse nerve net under the framework
Network can be used for the study and identification of handwriting digital pattern pulse sequence.When the spatiotemporal mode information that pattern pulse sequence is included
After becoming more complicated and being various, if it is desired to reach preferable pulse train recognition effect, then the volume of impulsive neural networks
Lamination, the number of plies of full articulamentum and scale require to be expanded, but pulse train proposed by the invention time for including and
More massive impulsive neural networks after the method that spatial information carries out combination learning is still suitable for expanding.
Present invention provide the technical scheme that
A kind of time of the pulse train based on Spike cube impulsive neural networks and spatial information carry out combination learning
Method, which is the coding mode information of two dimensional image, and each pixel of the sensitive chip of neuromorphic camera is most
The light intensity variation of detection itself in the time of small temporal resolution, when variable quantity is more than that itself corresponding excitation threshold is then sent out immediately
The signal of characterization a location of pixels and excitation instant out, and the signal is exported by bus, it is consequently formed and believes with the moment
The pulse train of breath and spatial coordinated information.Pulse train can be DVS, and (Dynamic Vision Sensor, dynamic vision pass
Sensor) camera, event mode camera, the pulse data for imitating the output of the neuromorphics cameras such as retina camera, software can also be passed through
Emulation generates, such as the pulse train generated after each pixel of still image is converted according to Poisson distribution.The present invention
Based on impulsive neural networks (spike cube and LIF neuron models) and STDP mechanism is used, to each of impulsive neural networks
The synapse weight and excitation threshold that layer neuron connects between each other are learnt, and learning process includes the following steps:
1) impulsive neural networks are built, the neuron of neural network is indicated using LIF model;
2) prepare pattern pulse sequence to be learned, be divided into training set according to certain ratio, verifying collects and surveys
Examination collection (can usually be divided according to the ratio of 5:1:1);
Spike cube (pulse train unit) is according to included in pattern pulse sequence to be learned or to be identified
The minimum time length of time series pattern carries out pulse train segment obtained from dividing at equal intervals to input picture pulse train.
The pattern pulse sequence of the input can be DVS (Dynamic Vision Sensor, dynamic visual sensor) camera, event
The pulse data of the neuromorphics camera outputs such as type camera, imitative retina camera, can also be generated by software emulation, such as
The pulse train that each pixel of still image generates after being converted according to Poisson distribution.
3) training set pattern pulse sequence inputting impulsive neural networks are subjected to propagated forward, pass sequentially through pulse nerve net
The first convolutional layer, the second convolutional layer, full articulamentum and the classification layer of network;In layer-by-layer communication process, simultaneously based on spike cube
The synapse weight and excitation threshold connected between each other using STDP mechanism to each layer neuron is learnt;It needs to illustrate
Be convolutional layer synapse weight and excitation threshold has different characteristics form at different times, to pattern pulse sequence when
Between and spatial information carry out combination learning method include: input calibration pattern pulse sequence, propagated forward and Layered Learning;
31) there will be the pattern pulse sequence inputting of calibration to impulsive neural networks;
In the study stage, input has the pattern pulse sequence of calibration, and the pattern pulse sequence is by a large amount of pulse train unit
(spike cube) composition;The acquisition modes of pulse train generally have 2 kinds of approach, the first is by DVS camera, event mode phase
Machine imitates the pulse train that the neuromorphics cameras such as retina camera are recorded some targets to be identified and exported, image
It is the minimum time length by these pulse trains according to pulse sequence mode needed for characterizing target to be identified that classification, which does calibration,
It is divided at equal intervals, the calibration value of each pulse train unit after division is neuromorphic camera corresponding to the segment
Record target.Second is to each pixel of still image by the characteristic of software analog neuron form camera according to Poisson point
Cloth is converted to pulse train unit to it, and the length of pulse train unit is set as pulse needed for characterizing target to be identified
The minimum time length of time series pattern, the calibration value of pulse train unit are set as its corresponding still image classification.32) forward direction
It propagates, comprising:
321) Pulse Calibration sequence propagated forward passes through the first convolutional layer;
The convolution kernel and its corresponding N number of characteristic pattern that first convolutional layer is 1 by 1*N step-length form, and characteristic pattern is equivalent to
The response region being made of a certain number of neurons, neuron uses LIF neuron models, when cynapse has pulse input,
The film potential of the model increases the intensity of cynapse in the arrival time of pulse, if membrane potential of neurons is not made to reach excitation threshold
Value, then film potential can carry out exponential damping with the time, if membrane potential of neurons is made to have reached excitation threshold, film potential is rapid
It drops and a pulse is excited to be forwarded to the connected neuron of next layer;Convolution kernel is equivalent in the characteristic pattern being attached thereto
The shared cynapse of neuron, each convolution kernel are connected to each feature according to current time all input pulse respective positions
Figure is corresponding to it the neuron in region, and the film potential of these neurons obtains accordingly according to the case where convolution synaptic input pulse
Variation.In each time scale, neuron will check whether the film potential of oneself has been more than excitation threshold.If in characteristic pattern
Certain membrane potential of neurons have been more than excitation threshold, then one pulse of generation is excited not swash if being not above excitation threshold
Send out pulse.
322) pulse propagated forward passes through the second convolutional layer;
The pulse of each characteristic pattern neuron excitation of first convolutional layer is transmitted to the second convolutional layer.Second convolutional layer is by N*M
The convolution kernel and its corresponding M characteristic pattern composition that step-length is 2.Neuron uses LIF neuron models, and step-length is 2 can be
The size reduction of second convolutional layer characteristic pattern is the half of the first convolutional layer characteristic pattern size, in each characteristic pattern of the second convolutional layer
Neuron pass through convolution kernel cynapse and corresponding neuron in N number of characteristic pattern of the first convolutional layer is connected.When the first convolution
After neuron excitation in layer, the correspondence neuron of each characteristic pattern of the second convolutional layer, volume Two are transmitted to by convolution cynapse
After the neuron of lamination receives pulse, itself is adjusted according to the case where convolution synaptic input pulse being attached thereto accordingly
Film potential value, and detect the film potential value of itself in each time scale, if reaching excitation threshold, excite a pulse
And it is transmitted to the neuron that next layer is attached thereto, if not reaching i.e. excitation threshold, not excitation pulse.
323) pulse propagated forward passes through full articulamentum;
The pulse of each characteristic pattern neuron excitation of second convolutional layer is transmitted to full articulamentum.Full articulamentum is by K LIF nerve
Member composition, the value of K can be set as 1.5 times of the second convolutional layer neuronal quantity.If the second convolutional layer neuronal quantity is
L, then each neuron of full articulamentum is connected by L cynapse with each neuron of the second convolutional layer, is shared between two layers
L*K cynapse.After the neuron excitation of the second convolutional layer, pulse can be transmitted to the mind that full articulamentum is attached thereto by cynapse
Through member.The neuron of full articulamentum is transmitted in the cynapse that each time scale is connected according to oneself adjusts itself come the case where pulse
Film potential, and detect whether current film potential is more than excitation threshold, if it exceeds excitation threshold, then excite a pulse and pass
It is delivered to the neuron that next layer is attached thereto, if not reaching excitation threshold, not excitation pulse.
324) pulse propagated forward reaches classification layer
The pulse of the complete each neuron excitation of articulamentum is transmitted to classification layer.Classification layer is made of Y LIF neuron, Y's
Quantity of the value corresponding to target to be identified in pulse train.Each neuron of classification layer passes through K cynapse and full articulamentum
Each neuron is connected, therefore from full articulamentum to classification a total of K*Y cynapse of layer.When the neuron of full articulamentum excites
Afterwards, pulse can be transmitted to each neuron of classification layer by cynapse.Classify layer neuron can each time scale according to
The pulse situation that the whole cynapses transmitting being attached thereto adjusts the film potential of itself, and detects whether film potential is more than excitation threshold
Value, if it exceeds excitation threshold, then excitation pulse and being recorded in itself corresponding variable is saved, if do not surpassed
Excitation threshold is crossed, then not excitation pulse and is not recorded.
Learning method includes spike cube STDP mechanism (the unsupervised STDP based on spike cube in the present invention
Habit mechanism) and spike cube BP-STDP (spike cube STDP based back-propagation algorithm,
The back-propagation algorithm of synaptic plasticity is relied on based on pulse train unit and pulse time) mechanism.Spike cube STDP machine
System is used for the study of convolutional layer synapse weight, and spike cube BP-STDP mechanism is for full articulamentum and classification layer synapse weight
Study.Entire learning process is that layering carries out, and first carries out the synapse weight study of spike cube STDP mechanism,
After study, then carry out the synapse weight study of spike cube BP-STDP mechanism.
4) after every study a period of time, the learning process of cynapse is temporarily interrupted, and verifying is collected pattern pulse sequence inputting
To current impulsive neural networks, the recognition accuracy test of image category is carried out;
If 5) recognition accuracy reaches expected standard, terminates learning process and save current synapse weight and excitation threshold
Value;If not reaching expected standard, continue to learn on this basis, until reaching expected standard or entire training set arteries and veins
Sequence is rushed to learn to finish;
6) learn after to get arrive trained impulsive neural networks.
By above-mentioned impulsive neural networks (spike cube and the LIF neuron models) and use STDP mechanism of being based on to pulse
The impulsive neural networks that the method that the time of sequence and spatial information carry out combination learning is learnt are known applied to image
Not, include the following steps:
1) the pattern pulse sequence without calibration is inputted;
The pattern pulse sequence without calibration is inputted in identification process, compared to the pattern pulse sequence of calibration, difference is
It does not need to demarcate the image category characterized when dividing spike cube;
2) propagated forward;
Propagated forward in image sequence identification process is identical with step " propagated forward " in learning process.Knowing
Each layer excitation threshold that study finishes suitably is turned down (for example, each neuron excitation threshold that study finishes is existed during not
0.9 times is adjusted in identification process) so that final discrimination has faint promotion;
3) image classification identifies;Include:
When each spike cube pattern pulse sequence units fully enter impulsive neural networks and transmit arrival classification
After layer neuron, the excitation pulse number stored in each classification corresponding variable of layer neuron is counted;
Then the most classification neuron of selective exitation pulse number is as most active neuron, if there is two minds
Pulse number through member excitation is identical, then selects the accumulation highest neuron of film potential as most active neuron, finally, arteries and veins
The image recognition result for rushing neural network is image category representated by most active neuron.For example, one section of pulse train
The impulsive neural networks are input to, pulse eventually leads to classification layer neuron excitation by propagated forward, when the pulse train list
After member fully enters impulsive neural networks, the pulse number of each classification layer neuron excitation is counted, might as well assume No. 2 nerves
The pulse number of member excitation is most, then impulsive neural networks are 2 to the recognition result of the pulse train unit.
It is above-mentioned that the impulsive neural networks learnt using combination learning method provided by the invention are applied to figure
As identification classification results also show, combination learning methodology acquistion provided by the invention to impulsive neural networks can realize
Classification and Identification is carried out to pulse train.
Compared with prior art, the beneficial effects of the present invention are:
The invention proposes a kind of time for including to pattern pulse sequence based on spike cube impulsive neural networks and
The method that spatial information carries out combination learning, the impulsive neural networks for learning to terminate can realize the identification to pattern pulse sequence.
It to Sequence Learning and is identified, first right by convolutional neural networks compared to already existing scheme, such as by Recognition with Recurrent Neural Network
Then the convolutional neural networks that study terminates are converted into corresponding impulsive neural networks form by still image study right again
The method of pattern pulse recognition sequence, core of the invention innovative point are to propose one kind based on spike cube and LIF nerve
Meta-model and the method that combination learning is carried out to the time of pulse train and spatial information using STDP mechanism.The invention
Ground is realized based on spike cube and STDP mechanism to temporal timing information included in pulse train and spatially
Location information combination learning, learning the impulsive neural networks finished has the ability that identify to pulse train, recycles and instructs
The model perfected carries out image classification identification.The present invention provides new side for the study of pattern pulse sequence and Classification and Identification
Method.
Detailed description of the invention
Fig. 1 is to carry out the pulse train segment that Poisson is converted into static handwriting digital 0 in the embodiment of the present invention;
Wherein, (a) static handwriting digital 0;(b) the pulse train segment that figure (a) Poisson is converted into.
Fig. 2 is a kind of existing simple impulsive neural networks framework.
Fig. 3 is that pulse granting and the accumulative effect that the moment is respectively 1,3,7 are emulated in the embodiment of the present invention;
Wherein, effect is provided in pulse when (a) emulation moment is 1;(b) the pulse accumulation effect when emulation moment is 3,
(c) the pulse accumulation effect when emulation moment is 7.
Fig. 4 is LIF membrane potential of neurons characteristic schematic diagram;
t1 (1)、t2 (1)Indicate No. 1 cynapse of LIF neuron in t1、t2Reception is to pulse, t3 (2)、t4 (2)Indicate LIF nerve
First No. 2 cynapses are in t3、t4Reception is to pulse, tj (f)Indicate f LIF neuron in tjMoment excites 1 pulse.
Specific embodiment
With reference to the accompanying drawing, the present invention, the model of but do not limit the invention in any way are further described by embodiment
It encloses.
The invention proposes a kind of time for including to pulse train based on spike cube impulsive neural networks and spaces
The method that information carries out combination learning, and can be applied to the classification identification of image.Wherein, pulse train includes pattern pulse sequence
Column, voice sequence information, vibration sequence information etc..
It is provided by the invention based on spike cube and LIF neuron models and using STDP mechanism to pulse train when
Between and spatial information carry out combination learning method.Implementation process is: can build a set of pulse mind with reference to the framework of Fig. 2 first
Through network, the neuron of neural network selects LIF model;It is ready for pulse train to be learned, it will according to certain ratio
It is divided into training set, verifying collection and test set (can usually divide according to the ratio of 5:1:1), then training set pulse train
Input pulse neural network carries out propagated forward, passes sequentially through the first convolutional layer, the second convolution of impulsive neural networks shown in Fig. 2
Layer, full articulamentum and classification layer;In layer-by-layer communication process, based on spike cube and using STDP mechanism to each layer neuron
The synapse weight and excitation threshold connected between each other is learnt;After every study a period of time, cynapse is temporarily interrupted
Habit process, and verifying collection pulse train is input to current impulsive neural networks and carries out discrimination test, if discrimination reaches
To expected standard, then terminates learning process and save current synapse weight and excitation threshold, if not reaching expected standard,
Continue to learn on this basis, learns to finish until reaching expected standard or entire training set pulse train;It is tied in study
Shu Hou can carry out discrimination test to test set pulse train with the impulsive neural networks.
Spike cube is to carry out according to the minimum time length of pulse sequence mode to be identified to input pulse sequence
Pulse train segment obtained from dividing at equal intervals.The pulse train of the input can be DVS (Dynamic Vision
Sensor, dynamic visual sensor) camera, event mode camera, the umber of pulse for imitating the output of the neuromorphics cameras such as retina camera
According to can also be generated by software emulation, such as generated after each pixel of still image is converted according to Poisson distribution
Pulse train, generated as shown in Figure 1, illustrating and carrying out Poisson conversion to static handwriting digital 0 by software emulation
Pulse train segment.
The method that time to pulse train and spatial information carry out combination learning includes: input Pulse Calibration sequence, preceding
Pulse train can be identified to propagation and Layered Learning, the impulsive neural networks after study, the process packet of identification
Include: input is without Pulse Calibration sequence, propagated forward and Classification and Identification.Specifically comprise the following steps:
Learning process:
1. inputting the pulse train of calibration
Impulsive neural networks learning algorithm proposed by the present invention needs to input largely the pulse sequence for having calibration in the study stage
Column, the acquisition modes of pulse train generally have 2 kinds of approach, the first is by DVS camera, event mode camera, imitative retina phase
The neuromorphics such as machine camera is recorded to some targets to be identified and the pulse train that exports, by these pulse trains according to
The minimum time length of pulse sequence mode needed for characterizing target to be identified is divided at equal intervals, each pulse after division
The calibration value of sequence fragment is the recording target of neuromorphic camera corresponding to the segment.Second is to simulate mind by software
Characteristic through form camera is converted to pulse train segment, pulse sequence to it according to Poisson distribution to each pixel of still image
The length of column-slice section is set as the minimum time length for pulse sequence mode needed for characterizing target to be identified, pulse train
The calibration value of segment is set as still image corresponding to the segment.(generation and calibration process of pulse train for ease of understanding,
I is illustrated with an example: the pulse output sequence recorded with software analog neuron form camera to handwritten numeral 7, when emulation
Between unit be set as 1 microsecond, as shown in figure 3, pulse granting and the accumulative effect that the emulation moment is 1,3,7 are respectively shown, when imitative
It when the true moment is 1, illustrates under minimum time resolution ratio, the output situation of pulse, the pulse that the single moment provides is completely not
The still image handwritten numeral 7 being converted can be symbolized;When emulating the moment is 3, the pulse accumulated can be preliminary
The elementary contour for the still image handwritten numeral 7 being converted is embodied, but is influenced by ambient noise pulse, still can not
Symbolize handwritten numeral 7;When emulating the moment is 7, the pulse accumulated has been able to be recorded than more completely symbolizing
Still image handwritten numeral 7, therefore the length of pulse train segment is set as 7.)
Be ready to it is whole have Pulse Calibration sequence fragment after, pulse train segment is input to pulse nerve net one by one
Network.
2. propagated forward
(1) pulse propagated forward passes through the first convolutional layer
There is the pulse train of calibration according to the precedence of time scale successively the pulse input pulse mind at each moment
Through network, the first convolutional layer of impulsive neural networks is arrived first at.First convolutional layer by 1*N step-length be 1 convolution kernel and its
Corresponding N number of characteristic pattern composition, characteristic pattern are equivalent to the response region being made of a certain number of neurons, and neuron uses
LIF neuron models, (when cynapse has pulse input, film potential can be in the arteries and veins as shown in Figure 4 for the membrane potential characteristics of the model
The arrival time of punching increases the intensity of the cynapse, if membrane potential of neurons is not made to reach excitation threshold, film potential can be with
The time carry out exponential damping, if membrane potential of neurons is made to have reached excitation threshold, film potential rapid drawdown simultaneously excites an arteries and veins
Red switch passs next layer connected neuron), convolution kernel is equivalent to the shared of neuron in the characteristic pattern being attached thereto
Cynapse, each convolution kernel are connected to each characteristic pattern and are corresponding to it area according to current time all input pulse respective positions
The film potential of the neuron in domain, these neurons is changed accordingly according to the case where convolution synaptic input pulse (if volume
The excited intensity of product cynapse is greater than inhibition strength, then film potential increases corresponding amount, conversely, film potential reduction is corresponding to it
Amount).In each time scale, neuron will check whether the film potential of oneself has been more than excitation threshold.If in characteristic pattern
Certain membrane potential of neurons be more than excitation threshold, then excite generate a pulse, if being not above excitation threshold, no
Excitation pulse.
(2) pulse propagated forward passes through the second convolutional layer
The pulse of each characteristic pattern neuron excitation of first convolutional layer can be transmitted to the second convolutional layer.Second convolutional layer is by N*M
The convolution kernel and its corresponding M characteristic pattern composition that a step-length is 2.Neuron uses LIF neuron models, and step-length is 2 can be with
It is the half of the first convolutional layer characteristic pattern size, each characteristic pattern of the second convolutional layer the size reduction of the second convolutional layer characteristic pattern
In neuron pass through convolution kernel cynapse and corresponding neuron in N number of characteristic pattern of the first convolutional layer is connected.Work as the first volume
After neuron excitation in lamination, it is transmitted to the correspondence neuron of each characteristic pattern of the second convolutional layer by convolution cynapse, second
After the neuron of convolutional layer receives pulse, certainly according to adjustment accordingly the case where the convolution synaptic input pulse being attached thereto
The film potential value of body, and the film potential value of itself is detected in each time scale, if reaching excitation threshold, excite an arteries and veins
The neuron that next layer is attached thereto is rushed and is transmitted to, if not reaching excitation threshold, not excitation pulse.
(3) pulse propagated forward passes through full articulamentum
The pulse of each characteristic pattern neuron excitation of second convolutional layer can be transmitted to full articulamentum.Full articulamentum is by K LIF mind
It is formed through member, the value of K can be set as 1.5 times of the second convolutional layer neuronal quantity.If the second convolutional layer neuronal quantity
For L, then each neuron of full articulamentum is connected by L cynapse with each neuron of the second convolutional layer, between two layers altogether
There is L*K cynapse.After the neuron excitation of the second convolutional layer, pulse can be transmitted to full articulamentum by cynapse and is attached thereto
Neuron.The neuron of full articulamentum adjusts certainly the case where each time scale is transmitted according to the cynapse oneself connected come pulse
The film potential of body, and detect whether current film potential is more than excitation threshold, if it exceeds excitation threshold, then excite a pulse simultaneously
It is transmitted to the neuron that next layer is attached thereto, if not reaching excitation threshold, not excitation pulse.
(4) pulse propagated forward reaches classification layer
The pulse of the complete each neuron excitation of articulamentum can be transmitted to classification layer.Classification layer is made of Y LIF neuron, Y
Value correspond to the quantity of target to be identified in pulse train.Each neuron of classification layer passes through K cynapse and full articulamentum
Each neuron be connected, therefore from full articulamentum to classification a total of K*Y cynapse of layer.When the neuron of full articulamentum excites
Afterwards, pulse can be transmitted to each neuron of classification layer by cynapse.Classify layer neuron can each time scale according to
The pulse situation that the whole cynapses transmitting being attached thereto adjusts the film potential of itself, and detects whether film potential is more than excitation threshold
Value, if it exceeds excitation threshold, then excitation pulse and being recorded in itself corresponding variable is saved, if do not surpassed
Excitation threshold is crossed, then not excitation pulse and is not recorded.
3. Layered Learning
Learning algorithm proposed by the present invention is by unsupervised STDP study mechanism and has the BP-STDP (STDP- of supervision
Based back-propagation algorithm relies on the back-propagation algorithm of synaptic plasticity based on pulse time) machine
System collectively constitutes.Unsupervised STDP mechanism is used for the study of convolutional layer synapse weight, has the BP-STDP mechanism of supervision for complete
The study of articulamentum and classification layer synapse weight.Entire learning process is that layering carries out, and first carries out unsupervised learning
The synapse weight of STDP mechanism learns, and after study, then carries out the synapse weight study for having supervision BP-STDP mechanism.It connects down
Come, Layered Learning process specifically introduced:
(1) unsupervised STDP study
Unsupervised STDP mechanism is used for the study of convolutional layer synapse weight.Learning process is that layering carries out, right first
The synapse weight of first convolutional layer is learnt.By the process of propagated forward it is found that after input pulse passes through the first convolutional layer,
The partial nerve member of first convolutional layer can excite.Then a film potential is only selected most from the neuron that each characteristic pattern excites
High neuron carries out the synaptic plasticity study of STDP mechanism, and inhibit in other characteristic patterns of this layer simultaneously same position other
Excite the study of neuron.If occurring the neuron of excitation on the same position of multiple characteristic patterns, film potential is only enabled
Highest learning of neuron, other neurons inhibit its learning process.If there is characteristic pattern in the highest nerve of film potential
Member is suppressed study, then the excitation neuron high to film potential time carries out the study of synapse weight, swashs if film potential time is high
Member of going crazy also is suppressed, then and so on, the excitation neuron until having traversed this feature figure whole.If there is characteristic pattern
Neuron or the neuron of excitation is not excited all to be suppressed study, then this feature figure is in current time scale without prominent
Touch plastic inquiry learning.When the cost function of the convolution synapse weight of first layer converges within a certain range, then terminate first layer
Study, and start to learn the synapse weight of the second convolutional layer with identical method.Second convolutional layer synapse weight
It is similar to the first convolutional layer to practise the judgment method terminated.
For the excitation threshold of convolutional layer neuron, it is also desirable to be learnt.Convolution synapse weight generally can be according to normal state
Distribution is initialized (mean value μ, variance δ), and excitation threshold is then generally initialized as the equal of convolution synapse weight initialization
Value μ.After convolution cynapse learns one section of long period interval, excitation threshold needs the learning effect according to convolution synapse weight
It is adjusted.The algorithm proposed in the present invention is to be set as neuron excitation threshold to be attached thereto being averaged for convolution synapse weight
Value is multiplied by a zoom factor, and the range of zoom factor is generally between 0.5~1.0.The learning process of excitation threshold is adjoint
The termination of convolution synapse weight learning process and terminate.
It is that study is single with spike cube pulse train segment in convolution weight and the learning process of excitation threshold
Position, it is assumed that spike cube pulse train fragment length is 5 time scales, then convolution weight and excitation threshold are in 5 times
There is different characteristic morphologies under scale respectively.The convolution weight at each current time will inherit last moment convolution weight letter
On the basis of breath, then carry out the plastic inquiry learning of weight at moment instantly.This but also the convolution cynapse that finishes of study different
Time scale carries out convolution with pulse of the different characteristic morphologies to input, goes to judge with the excitation threshold of different moments after convolution
Whether the neuron for having changed film potential should excite.It is associated with study by convolution cynapse and excitation threshold, makes convolution cynapse
The temporal information of pulse train is extracted.In addition, convolution is carried out to input pulse by multiple convolution, by input pulse sequence
The spatial model information for being included is mapped to higher-dimension from low-dimensional, carries out classification using high dimensional feature for classification layer and provides condition.
It goes to excavate the timing information for including in pulse train and space from the root in conclusion learning algorithm proposed by the present invention has
The ability of information.
(2) there is supervision BP-STDP study
There is study of the supervision BP-STDP study mechanism for full articulamentum and layer synapse weight of classifying.The process of study
It is that layering carries out, but first carries out the backward learning for having supervision since the last layer classifies layer.The number of classification layer neuron
Identical as the number of target to be identified in pulse train, therefore, each neuron for layer of classifying both corresponds to a mesh to be identified
Mark.In the learning process of synapse weight, the pulse train segment of input has calibration, and calibration value is the pulse train piece
The classification of the characterized target of section.When pulse propagated forward reach classification layer neuron after, the classification layer neuron of excitation then with this
The calibration value of pulse train is compared, and comparison result is artificially divided into 3 kinds of situations: the first is to represent current calibration value
The classification layer neuron of classification does not excite, and such case is marked as 1;Second is that the classification layer neuron excited is not generation
The current calibration value classification of table, such case are marked as -1;The third situation covers 2 kinds as a result, a kind of the result is that representing current
The classification neuron of calibration value classification excites, and another kind is not the result is that the neuron for not representing current calibration value classification swashs
Hair, the third situation are marked as 0.The learning algorithm proposed according to the present invention, for the first above-mentioned situation, excitation nerve
Synapse weight between member and full articulamentum is done according to STDP mechanism to be updated;For above-mentioned second situation, the nerve of the excitation
Synapse weight between member and full articulamentum will be done according to anti-STDP mechanism updates (anti-STDP mechanism and STDP Mechanism Primary
Reason is consistent, but symbol multiplies on the contrary, namely on the basis of STDP mechanism is to synapse weight knots modification in weight renewal process
With mark value -1).
In each spike cube study unit, whenever the study for having carried out a subseries layer synapse weight, according to anti-
To the principle of propagation, the study of primary full articulamentum synapse weight will be and then carried out.It is reversely passed with numerical value neural network
It broadcasts unlike algorithm, in numerical value neural network, backpropagation is carried out based on gradient, and in impulsive neural networks,
Backpropagation is the pulse situation based on neuron excitation.The synapse weight between the second convolutional layer of full articulamentum is specifically learned
Habit process is as follows: the excitation situation of each neuron of full articulamentum is traversed one by one first, if excited in time scale instantly
Pulse, then update the synapse weight being connected between the neuron and the second convolutional layer, the knots modification of weight and the neuron with
The excitation situation of the classification each neuron of layer is related with connection weight, i.e., the excitation situation (1, -1,0) of each classification layer neuron
Multiplied by the synapse weight respectively between the full articulamentum neuron and operation result is summed, this result is learned multiplied by one
Habit rate parameter and front convolutional layer correspond to the excited state of neuron, and (excitation does not excite for 1, is exactly for 0) obtained value
The knots modification of cynapse is connected between neuron neuron corresponding with front convolutional layer.According to the method described above, it updates and connects entirely one by one
The synapse weight that each neuron of layer is connected with front convolutional layer is connect, then completes primary full articulamentum synapse weight
It practises.It should be strongly noted that the neuron excitation threshold of full articulamentum does not need study, during initialization in advance
It is set as a fixed value.
Whenever the synapse weight of full articulamentum and classification layer temporarily ceases after the study of longer period of time
Then verifying collection pulse train is input to current impulsive neural networks and carries out discrimination test, if discrimination by habit process
Reach expected standard, then terminate learning process and saves synapse weight and excitation threshold, if not reaching expected standard,
Continue to learn on the basis of this, until reaching expected standard or after entire training set pulse train learns, end connects entirely
Connect the BP-STDP learning process of layer and layer neuronal synapse weight of classifying.
Identification process:
Impulsive neural networks after study can identify that the process of identification includes: input nothing to pulse train
Pulse Calibration sequence, propagated forward and Classification and Identification.Wherein, first step " pulse train of the input without calibration " and learning process
The first step in the pulse train of calibration " input " it is similar, do not need only to demarcate when dividing spike cube;
Second step " propagated forward " is identical with second step " propagated forward " in learning process, only in identification process
(for example, each neuron excitation threshold that study finishes was being identified after each layer excitation threshold that study finishes suitably is turned down
0.9 times is adjusted in journey) there can be faint promotion to final discrimination;Third step " Classification and Identification ", as each spike
After cube pulse train segment fully enters impulsive neural networks and transmits arrival classification layer neuron, each point is counted at this time
The excitation pulse number stored in the corresponding variable of class layer neuron, then the most classification of selective exitation pulse number is refreshing
It is used as most active neuron through member, it is identical if there is the pulse number of two neurons excitation, then select accumulation film potential
Highest neuron is as most active neuron, and finally, the classification results of impulsive neural networks output are most active nerves
Target category representated by member.
It should be noted that the purpose for publicizing and implementing example is to help to further understand the present invention, but the skill of this field
Art personnel, which are understood that, not to be departed from the present invention and spirit and scope of the appended claims, and various substitutions and modifications are all
It is possible.Therefore, the present invention should not be limited to embodiment disclosure of that, and the scope of protection of present invention is with claim
Subject to the range that book defines.
Claims (8)
1. a kind of method that the space time information for pattern pulse sequence carries out combination learning, establishes arteries and veins based on impulsive neural networks
Sequence units Spike cube and LIF neuron models are rushed, and use STDP mechanism, to each layer neuron of impulsive neural networks
The synapse weight and excitation threshold connected between each other is learnt;Include the following steps:
1) impulsive neural networks are established, the neuron of neural network is indicated using LIF model;
2) pattern pulse sequence to be learned is divided into training set, verifying collection and test set by ratio;
According to the minimum time length for the time series pattern for including in pattern pulse sequence to be learned, to the pattern pulse sequence of input
Column are divided at equal intervals obtains pulse train segment, and obtained pulse train segment is known as pulse train cell S pike
cube;
3) impulsive neural networks for establishing training set pattern pulse sequence inputting step 1) carry out propagated forward, pass sequentially through arteries and veins
Rush the first convolutional layer, the second convolutional layer, full articulamentum and the classification layer of neural network;
In layer-by-layer communication process, it is layered the synapse weight connected between each other to each layer neuron and excitation threshold
It practises, i.e. the time to pattern pulse sequence and spatial information carries out combination learning;
Combination learning method includes: pattern pulse sequence, propagated forward and the Layered Learning of input calibration;First it is based on pulse train
The unsupervised STDP study mechanism of unit carries out convolutional layer synapse weight and excitation threshold study, then based on pulse train unit and
The backpropagation mechanism that pulse time relies on synaptic plasticity carries out full articulamentum and classification layer synapse weight study;
4) after every study a period of time, the learning process of cynapse is temporarily interrupted, by verifying collection pattern pulse sequence inputting to currently
Impulsive neural networks, carry out image category recognition accuracy test;
If 5) recognition accuracy reaches expected standard, terminates learning process and save current synapse weight and excitation threshold;
If not reaching expected standard, continue to learn on this basis, until reaching expected standard or entire training set pulse
Sequence learns to finish;
6) learn after to get arrive trained impulsive neural networks.
2. the method for carrying out combination learning for the space time information of pattern pulse sequence as described in claim 1, characterized in that step
Rapid 3) combination learning method specifically comprises the following steps:
31) the pattern pulse sequence fragment for having calibration is input to impulsive neural networks one by one;
32) propagated forward and Layered Learning are carried out, comprising:
321) the pulse train segment propagated forward of calibration passes through the first convolutional layer;
First convolutional layer includes N number of convolution kernel and its corresponding N number of characteristic pattern, and characteristic pattern is the response area being made of neuron
Domain, neuron use LIF neuron models;Convolution kernel is the shared cynapse of neuron in the characteristic pattern being attached thereto, Mei Gejuan
Product core is connected to each characteristic pattern and is corresponding to it the neuron in region according to current time all input pulse respective positions;
The film potential of neuron is changed accordingly according to the case where convolution synaptic input pulse;
When cynapse has pulse input, film potential increases the intensity of cynapse in the arrival time of pulse;If not making neuron
Film potential reaches excitation threshold, then film potential can carry out exponential damping with the time, swashs if having reached membrane potential of neurons
Threshold value is sent out, then one pulse of film potential rapid drawdown and excitation is forwarded to the connected neuron of next layer;
In each time scale, neuron checks whether film potential is more than threshold value;If certain neuron membranes electricity in characteristic pattern
Position is more than threshold value, then excites and generate a pulse, if it does not exceed the threshold, then not excitation pulse;
322) pulse propagated forward passes through the second convolutional layer;
Second convolutional layer includes N*M convolution kernel and its corresponding M characteristic pattern;Neuron uses LIF neuron models;Second
The neuronal quantity of convolutional layer is L;After the neuron in the first convolutional layer excites, the second convolution is transmitted to by convolution cynapse
The correspondence neuron of each characteristic pattern of layer, after the neuron of the second convolutional layer receives pulse, according to the convolution being attached thereto
The case where synaptic input pulse, adjusts the film potential value of itself accordingly, and detects the film potential of itself in each time scale
Value excites a pulse and is transmitted to the neuron that next layer is attached thereto, if do not reached if reaching excitation threshold
Threshold value, then not excitation pulse;
323) pulse propagated forward passes through full articulamentum;
The pulse of each characteristic pattern neuron excitation of second convolutional layer is transmitted to full articulamentum;Full articulamentum includes K LIF nerve
Member;Then each neuron of full articulamentum is connected by L cynapse with each neuron of the second convolutional layer, is shared between two layers
L*K cynapse;
After the neuron excitation of the second convolutional layer, pulse is transmitted to the neuron that full articulamentum is attached thereto by cynapse;
The neuron of full articulamentum each time scale according to the cynapse of connection transmit come pulse adjust itself film potential,
And detect whether current film potential is more than threshold value;If it exceeds the threshold, then exciting a pulse and being transmitted to what next layer was connected
Neuron;If not reaching threshold value, not excitation pulse;
324) pulse propagated forward reaches classification layer;
The pulse of the complete each neuron excitation of articulamentum is transmitted to classification layer;Classification layer is made of Y LIF neuron, the value pair of Y
Should in pulse train target to be identified quantity;Each neuron of classification layer passes through each of K cynapse and full articulamentum
Neuron is connected, from full articulamentum to classification a total of K*Y cynapse of layer;
After the neuron excitation of full articulamentum, pulse is transmitted to each neuron of classification layer by cynapse;The mind of classification layer
Through member each time scale according to connected whole cynapses transmitting come pulse adjust the film potential of itself, and detect film potential
It whether is more than threshold value;
If it exceeds the threshold, then excitation pulse and being recorded in itself corresponding variable is saved, if be not above
Threshold value, then not excitation pulse.
3. the method for carrying out combination learning for the space time information of pattern pulse sequence as claimed in claim 2, characterized in that the
The convolution kernel of one convolutional layer is 1*N, and convolution kernel step-length is 1;The convolution kernel of second convolutional layer is N*M, step-length 2;By
The size reduction of two convolutional layer characteristic patterns is the half of the first convolutional layer characteristic pattern size.
4. the method for carrying out combination learning for the space time information of pattern pulse sequence as described in claim 1, characterized in that have
The acquisition of the pattern pulse sequence of calibration includes two methods:
For carrying out recording the pulse train exported to target by neuromorphic camera, specific method is: by pulse train
It is divided at equal intervals according to the minimum time length of pulse sequence mode needed for characterization target, each pulse sequence after division
The calibration value of column-slice section is the recording target of neuromorphic camera corresponding to the segment;
The pulse train segment that characteristic for emulating neuromorphic camera by software tool generates, specific method is: by arteries and veins
The length for rushing sequence fragment is set as minimum time length for pulse sequence mode needed for characterizing target to be identified, pulse
The calibration value of sequence fragment is set as still image corresponding to the segment.
5. the method for carrying out combination learning for the space time information of pattern pulse sequence as described in Claims 1 to 4, feature
It is that pulse train can also be voice sequence information or vibration sequence information.
6. a kind of image-recognizing method for carrying out combination learning based on pattern pulse sequence space time information, utilizes Claims 1 to 44
The trained impulsive neural networks that the method that the space time information for pattern pulse sequence carries out combination learning obtains into
The identification of row image classification, includes the following steps:
S1 the pattern pulse sequence without calibration) is inputted;Without being carried out when dividing Spike cube to the image category characterized
Calibration;
S2 the step of) executing the step 3) propagated forward, carries out propagated forward;
S3) Classification and Identification;Include:
S31) each Spike cube pattern pulse sequence fragment fully enters impulsive neural networks and transmits arrival classification layer
Neuron;
S32 after), the excitation pulse number stored in each classification corresponding variable of layer neuron is counted;
S33) then using the largest number of classification neurons of excitation pulse as most active neuron;
S34) if the pulse number of two neurons excitation is identical, it is most active that the highest neuron conduct of film potential will be accumulated
Neuron;
S35 image category representated by most active neuron) is exported, as image classification result.
7. the image-recognizing method of combination learning is carried out based on pattern pulse sequence space time information as claimed in claim 6, it is special
Sign is step S2) value that each neuron excitation threshold that study finishes is turned down in propagated forward is carried out, to promote discrimination.
8. the image-recognizing method of combination learning is carried out based on pattern pulse sequence space time information as claimed in claim 6, it is special
Sign is that pattern pulse sequence is the pulse data of neuromorphic camera output or the pulse train by emulation generation;The mind
Include but is not limited to through form camera: dynamic visual sensor camera, event mode camera, imitative retina camera.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910481420.9A CN110210563B (en) | 2019-06-04 | 2019-06-04 | Image pulse data space-time information learning and identification method based on Spike cube SNN |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910481420.9A CN110210563B (en) | 2019-06-04 | 2019-06-04 | Image pulse data space-time information learning and identification method based on Spike cube SNN |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110210563A true CN110210563A (en) | 2019-09-06 |
CN110210563B CN110210563B (en) | 2021-04-30 |
Family
ID=67790543
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910481420.9A Active CN110210563B (en) | 2019-06-04 | 2019-06-04 | Image pulse data space-time information learning and identification method based on Spike cube SNN |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110210563B (en) |
Cited By (47)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110781968A (en) * | 2019-10-29 | 2020-02-11 | 中国人民解放军国防科技大学 | Extensible class image identification method based on plastic convolution neural network |
CN110826602A (en) * | 2019-10-23 | 2020-02-21 | 中国科学院自动化研究所 | Image classification method and system based on membrane potential regulation and control pulse neural network |
CN110826437A (en) * | 2019-10-23 | 2020-02-21 | 中国科学院自动化研究所 | Intelligent robot control method, system and device based on biological neural network |
CN110991610A (en) * | 2019-11-28 | 2020-04-10 | 华中科技大学 | Probabilistic neuron circuit, probabilistic neural network topological structure and application thereof |
CN111046954A (en) * | 2019-12-12 | 2020-04-21 | 电子科技大学 | Image classification method of spiking learning model based on dynamic threshold |
CN111091815A (en) * | 2019-12-12 | 2020-05-01 | 电子科技大学 | Voice recognition method of aggregation label learning model based on membrane voltage driving |
CN111275742A (en) * | 2020-01-19 | 2020-06-12 | 北京大学 | Target identification method, device and system and computer readable storage medium |
CN111340194A (en) * | 2020-03-02 | 2020-06-26 | 中国科学技术大学 | Pulse convolution neural network neural morphology hardware and image identification method thereof |
CN111488908A (en) * | 2020-03-10 | 2020-08-04 | 天津大学 | Brain-imitating image identification method based on enhanced pulse |
CN111612136A (en) * | 2020-05-25 | 2020-09-01 | 之江实验室 | Neural morphology visual target classification method and system |
CN112085768A (en) * | 2020-09-02 | 2020-12-15 | 北京灵汐科技有限公司 | Optical flow information prediction method, optical flow information prediction device, electronic device, and storage medium |
CN112130118A (en) * | 2020-08-19 | 2020-12-25 | 复旦大学无锡研究院 | SNN-based ultra-wideband radar signal processing system and processing method |
CN112155549A (en) * | 2020-09-04 | 2021-01-01 | 西北师范大学 | ADHD disease diagnosis aid decision-making system based on deep convolution pulse neural network |
CN112418296A (en) * | 2020-11-18 | 2021-02-26 | 中国科学院上海微系统与信息技术研究所 | Bionic binocular target recognition and tracking method based on human eye visual attention mechanism |
CN112633497A (en) * | 2020-12-21 | 2021-04-09 | 中山大学 | Convolutional pulse neural network training method based on reweighted membrane voltage |
CN112699956A (en) * | 2021-01-08 | 2021-04-23 | 西安交通大学 | Neural morphology visual target classification method based on improved impulse neural network |
CN112767501A (en) * | 2021-01-06 | 2021-05-07 | 西南大学 | VCSEL-SA image identification system and method based on electric control stimulation |
CN112762100A (en) * | 2021-01-14 | 2021-05-07 | 哈尔滨理工大学 | Bearing full-life-cycle monitoring method based on digital twinning |
CN112906828A (en) * | 2021-04-08 | 2021-06-04 | 周士博 | Image classification method based on time domain coding and impulse neural network |
CN113077017A (en) * | 2021-05-24 | 2021-07-06 | 河南大学 | Synthetic aperture image classification method based on impulse neural network |
CN113095492A (en) * | 2021-04-14 | 2021-07-09 | 北京大学 | Topological feature detection method and device based on biological neural network |
CN113111758A (en) * | 2021-04-06 | 2021-07-13 | 中山大学 | SAR image ship target identification method based on pulse neural network |
CN113255905A (en) * | 2021-07-16 | 2021-08-13 | 成都时识科技有限公司 | Signal processing method of neurons in impulse neural network and network training method |
CN113269264A (en) * | 2021-06-04 | 2021-08-17 | 北京灵汐科技有限公司 | Object recognition method, electronic device, and computer-readable medium |
CN113269313A (en) * | 2021-06-04 | 2021-08-17 | 北京灵汐科技有限公司 | Synapse weight training method, electronic device and computer readable medium |
CN113269113A (en) * | 2021-06-04 | 2021-08-17 | 北京灵汐科技有限公司 | Human behavior recognition method, electronic device, and computer-readable medium |
CN113375676A (en) * | 2021-05-26 | 2021-09-10 | 南京航空航天大学 | Detector landing point positioning method based on impulse neural network |
CN113408613A (en) * | 2021-06-18 | 2021-09-17 | 电子科技大学 | Single-layer image classification method based on delay mechanism |
CN113397482A (en) * | 2021-05-19 | 2021-09-17 | 中国航天科工集团第二研究院 | Human behavior analysis method and system |
WO2021233179A1 (en) * | 2020-05-19 | 2021-11-25 | 深圳忆海原识科技有限公司 | Brain-like visual neural network having forward learning and meta-learning functions |
CN113902106A (en) * | 2021-12-06 | 2022-01-07 | 成都时识科技有限公司 | Pulse event decision device, method, chip and electronic equipment |
CN113962371A (en) * | 2021-12-23 | 2022-01-21 | 中科南京智能技术研究院 | Image identification method and system based on brain-like computing platform |
CN114022652A (en) * | 2020-07-15 | 2022-02-08 | 中移(苏州)软件技术有限公司 | Data processing method, equipment and device and computer storage medium |
CN114037050A (en) * | 2021-10-21 | 2022-02-11 | 大连理工大学 | Robot degradation environment obstacle avoidance method based on internal plasticity of pulse neural network |
CN114065806A (en) * | 2021-10-28 | 2022-02-18 | 贵州大学 | Manipulator touch data classification method based on impulse neural network |
CN114091663A (en) * | 2021-11-28 | 2022-02-25 | 重庆大学 | Lightweight on-chip learning method, system and processor based on impulse neural network |
CN114202068A (en) * | 2022-02-17 | 2022-03-18 | 浙江大学 | Self-learning implementation system for brain-like computing chip |
CN114220089A (en) * | 2021-11-29 | 2022-03-22 | 北京理工大学 | Method for carrying out pattern recognition based on segmented progressive pulse neural network |
CN114429491A (en) * | 2022-04-07 | 2022-05-03 | 之江实验室 | Pulse neural network target tracking method and system based on event camera |
CN114659553A (en) * | 2022-02-28 | 2022-06-24 | 联想(北京)有限公司 | Detection method, device, equipment and storage medium |
CN114881070A (en) * | 2022-04-07 | 2022-08-09 | 河北工业大学 | AER object identification method based on bionic hierarchical pulse neural network |
CN115238857A (en) * | 2022-06-15 | 2022-10-25 | 脉冲视觉(北京)科技有限公司 | Neural network based on pulse signal and pulse signal processing method |
CN115246559A (en) * | 2022-07-22 | 2022-10-28 | 鄂尔多斯市国源矿业开发有限责任公司 | Industrial belt longitudinal tearing identification method |
CN115810138A (en) * | 2022-11-18 | 2023-03-17 | 天津大学 | Image identification method based on multi-electrode array in-vitro culture neuron network |
CN115994563A (en) * | 2022-10-31 | 2023-04-21 | 天津大学 | Brain-like situation learning model construction and training method for intelligent auxiliary driving |
WO2023151289A1 (en) * | 2022-02-09 | 2023-08-17 | 苏州浪潮智能科技有限公司 | Emotion identification method, training method, apparatus, device, storage medium and product |
US11954579B2 (en) | 2021-06-04 | 2024-04-09 | Lynxi Technologies Co., Ltd. | Synaptic weight training method, target identification method, electronic device and medium |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105404902A (en) * | 2015-10-27 | 2016-03-16 | 清华大学 | Impulsive neural network-based image feature describing and memorizing method |
US20170236027A1 (en) * | 2016-02-16 | 2017-08-17 | Brainchip Inc. | Intelligent biomorphic system for pattern recognition with autonomous visual feature extraction |
EP3340121A1 (en) * | 2016-12-20 | 2018-06-27 | Intel Corporation | Network traversal using neuromorphic instantiations of spike-time-dependent plasticity |
CN108875846A (en) * | 2018-05-08 | 2018-11-23 | 河海大学常州校区 | A kind of Handwritten Digit Recognition method based on improved impulsive neural networks |
CN109102000A (en) * | 2018-09-05 | 2018-12-28 | 杭州电子科技大学 | A kind of image-recognizing method extracted based on layered characteristic with multilayer impulsive neural networks |
CN109635938A (en) * | 2018-12-29 | 2019-04-16 | 电子科技大学 | A kind of autonomous learning impulsive neural networks weight quantization method |
-
2019
- 2019-06-04 CN CN201910481420.9A patent/CN110210563B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105404902A (en) * | 2015-10-27 | 2016-03-16 | 清华大学 | Impulsive neural network-based image feature describing and memorizing method |
US20170236027A1 (en) * | 2016-02-16 | 2017-08-17 | Brainchip Inc. | Intelligent biomorphic system for pattern recognition with autonomous visual feature extraction |
EP3340121A1 (en) * | 2016-12-20 | 2018-06-27 | Intel Corporation | Network traversal using neuromorphic instantiations of spike-time-dependent plasticity |
CN108875846A (en) * | 2018-05-08 | 2018-11-23 | 河海大学常州校区 | A kind of Handwritten Digit Recognition method based on improved impulsive neural networks |
CN109102000A (en) * | 2018-09-05 | 2018-12-28 | 杭州电子科技大学 | A kind of image-recognizing method extracted based on layered characteristic with multilayer impulsive neural networks |
CN109635938A (en) * | 2018-12-29 | 2019-04-16 | 电子科技大学 | A kind of autonomous learning impulsive neural networks weight quantization method |
Non-Patent Citations (3)
Title |
---|
AMIRHOSSEIN TAVANAEI ET AL: "Multi-Layer Unsupervised Learning in a Spiking Convolutional Neural Network", 《2017 INTERNATIONAL JOINT CONFERENCE ON NEURAL NETWORKS (IJCNN)》 * |
WEI WANG ET AL: "Learning of spatiotemporal patterns in a spiking neural network with resistive switching synapses", 《SCIENCE ADVANCES》 * |
蔺想红等: "基于脉冲序列核的脉冲神经元监督学习算法", 《电子学报》 * |
Cited By (75)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110826602A (en) * | 2019-10-23 | 2020-02-21 | 中国科学院自动化研究所 | Image classification method and system based on membrane potential regulation and control pulse neural network |
CN110826437A (en) * | 2019-10-23 | 2020-02-21 | 中国科学院自动化研究所 | Intelligent robot control method, system and device based on biological neural network |
CN110826602B (en) * | 2019-10-23 | 2022-04-26 | 中国科学院自动化研究所 | Image classification method and system based on membrane potential regulation and control pulse neural network |
CN110781968A (en) * | 2019-10-29 | 2020-02-11 | 中国人民解放军国防科技大学 | Extensible class image identification method based on plastic convolution neural network |
CN110991610B (en) * | 2019-11-28 | 2022-08-05 | 华中科技大学 | Probability determination method for nondeterministic problem |
CN110991610A (en) * | 2019-11-28 | 2020-04-10 | 华中科技大学 | Probabilistic neuron circuit, probabilistic neural network topological structure and application thereof |
CN111091815A (en) * | 2019-12-12 | 2020-05-01 | 电子科技大学 | Voice recognition method of aggregation label learning model based on membrane voltage driving |
CN111046954A (en) * | 2019-12-12 | 2020-04-21 | 电子科技大学 | Image classification method of spiking learning model based on dynamic threshold |
CN111275742A (en) * | 2020-01-19 | 2020-06-12 | 北京大学 | Target identification method, device and system and computer readable storage medium |
WO2021143066A1 (en) * | 2020-01-19 | 2021-07-22 | 北京大学 | Target recognition method, device, and system, and computer readable storage medium |
CN111275742B (en) * | 2020-01-19 | 2022-01-11 | 北京大学 | Target identification method, device and system and computer readable storage medium |
CN111340194A (en) * | 2020-03-02 | 2020-06-26 | 中国科学技术大学 | Pulse convolution neural network neural morphology hardware and image identification method thereof |
CN111340194B (en) * | 2020-03-02 | 2022-09-06 | 中国科学技术大学 | Pulse convolution neural network neural morphology hardware and image identification method thereof |
CN111488908A (en) * | 2020-03-10 | 2020-08-04 | 天津大学 | Brain-imitating image identification method based on enhanced pulse |
WO2021233179A1 (en) * | 2020-05-19 | 2021-11-25 | 深圳忆海原识科技有限公司 | Brain-like visual neural network having forward learning and meta-learning functions |
CN111612136B (en) * | 2020-05-25 | 2023-04-07 | 之江实验室 | Neural morphology visual target classification method and system |
CN111612136A (en) * | 2020-05-25 | 2020-09-01 | 之江实验室 | Neural morphology visual target classification method and system |
CN114022652A (en) * | 2020-07-15 | 2022-02-08 | 中移(苏州)软件技术有限公司 | Data processing method, equipment and device and computer storage medium |
CN112130118B (en) * | 2020-08-19 | 2023-11-17 | 复旦大学无锡研究院 | Ultra-wideband radar signal processing system and method based on SNN |
CN112130118A (en) * | 2020-08-19 | 2020-12-25 | 复旦大学无锡研究院 | SNN-based ultra-wideband radar signal processing system and processing method |
CN112085768B (en) * | 2020-09-02 | 2023-12-26 | 北京灵汐科技有限公司 | Optical flow information prediction method, optical flow information prediction device, electronic equipment and storage medium |
CN112085768A (en) * | 2020-09-02 | 2020-12-15 | 北京灵汐科技有限公司 | Optical flow information prediction method, optical flow information prediction device, electronic device, and storage medium |
CN112155549A (en) * | 2020-09-04 | 2021-01-01 | 西北师范大学 | ADHD disease diagnosis aid decision-making system based on deep convolution pulse neural network |
CN112155549B (en) * | 2020-09-04 | 2023-11-14 | 西北师范大学 | ADHD disease diagnosis auxiliary decision-making system based on deep convolution impulse neural network |
CN112418296A (en) * | 2020-11-18 | 2021-02-26 | 中国科学院上海微系统与信息技术研究所 | Bionic binocular target recognition and tracking method based on human eye visual attention mechanism |
CN112418296B (en) * | 2020-11-18 | 2024-04-02 | 中国科学院上海微系统与信息技术研究所 | Bionic binocular target identification and tracking method based on human eye visual attention mechanism |
CN112633497A (en) * | 2020-12-21 | 2021-04-09 | 中山大学 | Convolutional pulse neural network training method based on reweighted membrane voltage |
CN112633497B (en) * | 2020-12-21 | 2023-08-18 | 中山大学 | Convolutional impulse neural network training method based on re-weighted membrane voltage |
CN112767501A (en) * | 2021-01-06 | 2021-05-07 | 西南大学 | VCSEL-SA image identification system and method based on electric control stimulation |
CN112699956B (en) * | 2021-01-08 | 2023-09-22 | 西安交通大学 | Neuromorphic visual target classification method based on improved impulse neural network |
CN112699956A (en) * | 2021-01-08 | 2021-04-23 | 西安交通大学 | Neural morphology visual target classification method based on improved impulse neural network |
CN112762100A (en) * | 2021-01-14 | 2021-05-07 | 哈尔滨理工大学 | Bearing full-life-cycle monitoring method based on digital twinning |
CN113111758A (en) * | 2021-04-06 | 2021-07-13 | 中山大学 | SAR image ship target identification method based on pulse neural network |
CN113111758B (en) * | 2021-04-06 | 2024-01-12 | 中山大学 | SAR image ship target recognition method based on impulse neural network |
CN112906828A (en) * | 2021-04-08 | 2021-06-04 | 周士博 | Image classification method based on time domain coding and impulse neural network |
CN113095492A (en) * | 2021-04-14 | 2021-07-09 | 北京大学 | Topological feature detection method and device based on biological neural network |
CN113397482A (en) * | 2021-05-19 | 2021-09-17 | 中国航天科工集团第二研究院 | Human behavior analysis method and system |
CN113077017A (en) * | 2021-05-24 | 2021-07-06 | 河南大学 | Synthetic aperture image classification method based on impulse neural network |
CN113077017B (en) * | 2021-05-24 | 2022-12-13 | 河南大学 | Synthetic aperture image classification method based on pulse neural network |
CN113375676B (en) * | 2021-05-26 | 2024-02-20 | 南京航空航天大学 | Detector landing site positioning method based on impulse neural network |
CN113375676A (en) * | 2021-05-26 | 2021-09-10 | 南京航空航天大学 | Detector landing point positioning method based on impulse neural network |
US11954579B2 (en) | 2021-06-04 | 2024-04-09 | Lynxi Technologies Co., Ltd. | Synaptic weight training method, target identification method, electronic device and medium |
CN113269113B (en) * | 2021-06-04 | 2024-04-30 | 北京灵汐科技有限公司 | Human behavior recognition method, electronic device, and computer-readable medium |
CN113269313B (en) * | 2021-06-04 | 2024-05-10 | 北京灵汐科技有限公司 | Synaptic weight training method, electronic device, and computer-readable medium |
CN113269264B (en) * | 2021-06-04 | 2024-07-26 | 北京灵汐科技有限公司 | Target recognition method, electronic device, and computer-readable medium |
CN113269113A (en) * | 2021-06-04 | 2021-08-17 | 北京灵汐科技有限公司 | Human behavior recognition method, electronic device, and computer-readable medium |
CN113269313A (en) * | 2021-06-04 | 2021-08-17 | 北京灵汐科技有限公司 | Synapse weight training method, electronic device and computer readable medium |
CN113269264A (en) * | 2021-06-04 | 2021-08-17 | 北京灵汐科技有限公司 | Object recognition method, electronic device, and computer-readable medium |
CN113408613A (en) * | 2021-06-18 | 2021-09-17 | 电子科技大学 | Single-layer image classification method based on delay mechanism |
CN113408613B (en) * | 2021-06-18 | 2022-07-19 | 电子科技大学 | Single-layer image classification method based on delay mechanism |
CN113255905B (en) * | 2021-07-16 | 2021-11-02 | 成都时识科技有限公司 | Signal processing method of neurons in impulse neural network and network training method |
CN113255905A (en) * | 2021-07-16 | 2021-08-13 | 成都时识科技有限公司 | Signal processing method of neurons in impulse neural network and network training method |
CN114037050A (en) * | 2021-10-21 | 2022-02-11 | 大连理工大学 | Robot degradation environment obstacle avoidance method based on internal plasticity of pulse neural network |
CN114037050B (en) * | 2021-10-21 | 2022-08-16 | 大连理工大学 | Robot degradation environment obstacle avoidance method based on internal plasticity of pulse neural network |
CN114065806B (en) * | 2021-10-28 | 2022-12-20 | 贵州大学 | Manipulator touch data classification method based on impulse neural network |
CN114065806A (en) * | 2021-10-28 | 2022-02-18 | 贵州大学 | Manipulator touch data classification method based on impulse neural network |
CN114091663A (en) * | 2021-11-28 | 2022-02-25 | 重庆大学 | Lightweight on-chip learning method, system and processor based on impulse neural network |
CN114220089B (en) * | 2021-11-29 | 2024-06-14 | 北京理工大学 | Method for pattern recognition based on sectional progressive pulse neural network |
CN114220089A (en) * | 2021-11-29 | 2022-03-22 | 北京理工大学 | Method for carrying out pattern recognition based on segmented progressive pulse neural network |
CN113902106A (en) * | 2021-12-06 | 2022-01-07 | 成都时识科技有限公司 | Pulse event decision device, method, chip and electronic equipment |
CN113902106B (en) * | 2021-12-06 | 2022-02-22 | 成都时识科技有限公司 | Pulse event decision device, method, chip and electronic equipment |
CN113962371B (en) * | 2021-12-23 | 2022-05-20 | 中科南京智能技术研究院 | Image identification method and system based on brain-like computing platform |
CN113962371A (en) * | 2021-12-23 | 2022-01-21 | 中科南京智能技术研究院 | Image identification method and system based on brain-like computing platform |
WO2023151289A1 (en) * | 2022-02-09 | 2023-08-17 | 苏州浪潮智能科技有限公司 | Emotion identification method, training method, apparatus, device, storage medium and product |
CN114202068A (en) * | 2022-02-17 | 2022-03-18 | 浙江大学 | Self-learning implementation system for brain-like computing chip |
CN114202068B (en) * | 2022-02-17 | 2022-06-28 | 浙江大学 | Self-learning implementation system for brain-like computing chip |
CN114659553A (en) * | 2022-02-28 | 2022-06-24 | 联想(北京)有限公司 | Detection method, device, equipment and storage medium |
CN114429491A (en) * | 2022-04-07 | 2022-05-03 | 之江实验室 | Pulse neural network target tracking method and system based on event camera |
CN114881070A (en) * | 2022-04-07 | 2022-08-09 | 河北工业大学 | AER object identification method based on bionic hierarchical pulse neural network |
CN115238857B (en) * | 2022-06-15 | 2023-05-05 | 北京融合未来技术有限公司 | Neural network based on pulse signals and pulse signal processing method |
CN115238857A (en) * | 2022-06-15 | 2022-10-25 | 脉冲视觉(北京)科技有限公司 | Neural network based on pulse signal and pulse signal processing method |
CN115246559A (en) * | 2022-07-22 | 2022-10-28 | 鄂尔多斯市国源矿业开发有限责任公司 | Industrial belt longitudinal tearing identification method |
CN115994563B (en) * | 2022-10-31 | 2023-08-18 | 天津大学 | Brain-like situation learning model construction and training method for intelligent auxiliary driving |
CN115994563A (en) * | 2022-10-31 | 2023-04-21 | 天津大学 | Brain-like situation learning model construction and training method for intelligent auxiliary driving |
CN115810138A (en) * | 2022-11-18 | 2023-03-17 | 天津大学 | Image identification method based on multi-electrode array in-vitro culture neuron network |
Also Published As
Publication number | Publication date |
---|---|
CN110210563B (en) | 2021-04-30 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110210563A (en) | The study of pattern pulse data space time information and recognition methods based on Spike cube SNN | |
KR102641116B1 (en) | Method and device to recognize image and method and device to train recognition model based on data augmentation | |
CN107403154A (en) | A kind of gait recognition method based on dynamic visual sensor | |
CN107169435B (en) | Convolutional neural network human body action classification method based on radar simulation image | |
CN108629380B (en) | Cross-scene wireless signal sensing method based on transfer learning | |
CN104217214B (en) | RGB D personage's Activity recognition methods based on configurable convolutional neural networks | |
CN106982359B (en) | Binocular video monitoring method and system and computer readable storage medium | |
Wysoski et al. | Evolving spiking neural networks for audiovisual information processing | |
CN109583322A (en) | A kind of recognition of face depth network training method and system | |
CN106503642B (en) | A kind of model of vibration method for building up applied to optical fiber sensing system | |
CN105760930A (en) | Multilayer spiking neural network recognition system for AER | |
CN108520199A (en) | Based on radar image and the human action opener recognition methods for generating confrontation model | |
CN107862668A (en) | A kind of cultural relic images restored method based on GNN | |
CN108764050A (en) | Skeleton Activity recognition method, system and equipment based on angle independence | |
US20170337469A1 (en) | Anomaly detection using spiking neural networks | |
CN108960207A (en) | A kind of method of image recognition, system and associated component | |
KR20160138042A (en) | Invariant object representation of images using spiking neural networks | |
CN106875004A (en) | Composite mode neuronal messages processing method and system | |
CN106991666A (en) | A kind of disease geo-radar image recognition methods suitable for many size pictorial informations | |
CN114529484B (en) | Deep learning sample enhancement method for direct current component change in imaging | |
CN103700118B (en) | Based on the moving target detection method of pulse coupled neural network | |
CN108805879A (en) | A kind of image partition method based on Spiking neural networks | |
CN113553918B (en) | Machine ticket issuing character recognition method based on pulse active learning | |
CN103985115A (en) | Image multi-strength edge detection method having visual photosensitive layer simulation function | |
CN112862084B (en) | Traffic flow prediction method based on deep migration fusion learning |
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 |