CN111488908A - Brain-imitating image identification method based on enhanced pulse - Google Patents

Brain-imitating image identification method based on enhanced pulse Download PDF

Info

Publication number
CN111488908A
CN111488908A CN202010161401.0A CN202010161401A CN111488908A CN 111488908 A CN111488908 A CN 111488908A CN 202010161401 A CN202010161401 A CN 202010161401A CN 111488908 A CN111488908 A CN 111488908A
Authority
CN
China
Prior art keywords
pulse
enhanced
time
neuron
pulses
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.)
Pending
Application number
CN202010161401.0A
Other languages
Chinese (zh)
Inventor
于强
宋世明
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tianjin University
Original Assignee
Tianjin University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tianjin University filed Critical Tianjin University
Priority to CN202010161401.0A priority Critical patent/CN111488908A/en
Publication of CN111488908A publication Critical patent/CN111488908A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computational Linguistics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Evolutionary Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a brain-imitating image identification method based on enhanced pulses. Two new learning algorithms were first proposed to handle the enhancement pulse. Then, a plurality of enhanced pulse-based brain-like image recognition methods are provided by combining the current pulse coding method. The experimental result shows that compared with the prior model based on the binary form pulse, the image recognition technology based on the enhanced pulse greatly improves the performance of the SNN on the image recognition task, and highlights the advantages of the enhanced pulse and the learning method. It is noted that the enhanced pulse proposed in the present invention can be easily extended to other pulse-based systems, which is beneficial to further explore more efficient image recognition frameworks and provide new potential research directions for neuromorphic computation.

Description

Brain-imitating image identification method based on enhanced pulse
Technical Field
The invention belongs to the field of brain-like calculation and image recognition, in particular relates to a technology for improving the performance of a pulse neural network model on image recognition, and particularly relates to a brain-like image recognition method based on enhanced pulses.
Background
Neurons in biological systems use impulses to transmit and process information, in contrast to the analog values used in current deep neural networks. Therefore, a Spiking Neural Network (SNN) has been proposed to simulate and explore a pulse-based efficient and effective way of information processing in the mammalian brain. SNNs are considered to have better biological confidence and computational power due to having a similar information processing approach as biological systems.
However, all current image recognition techniques based on impulse neural networks are based on impulses in binary form, i.e. information is carried only by the exact impulse firing time. In the nervous system, a phenomenon of burst (burst) of pulses is ubiquitous, that is, a large number of pulses are emitted in a short time. We consider that this reveals another way of additional information transfer in addition to the burst length.
Disclosure of Invention
With this in mind, we introduce a new concept of enhancing pulses by defining a pulse coefficient to represent other information than the pulse firing time. However, it is unclear how neurons process and learn such enhancement pulses. In the present invention, we first propose two new learning algorithms to process the enhancement pulse. Then, a plurality of enhanced pulse-based brain-like image recognition methods are provided by combining the current pulse coding method. The experimental result shows that compared with the prior model based on the binary form pulse, the image recognition technology based on the enhanced pulse greatly improves the performance of the SNN on the image recognition task, and highlights the advantages of the enhanced pulse and the learning method. It is noted that the enhanced pulse proposed in the present invention can be easily extended to other pulse-based systems, which is beneficial to further explore more efficient image recognition frameworks and provide new potential research directions for neuromorphic computation.
The present invention first proposes a new concept of enhancing pulses. Thereafter, two new learning algorithms are proposed to process the enhancement pulses. By combining the current pulse coding method, a plurality of enhanced pulse-based brain-like image recognition methods are provided. The specific technical scheme is as follows: the brain-imitating image identification method based on the enhanced pulse comprises the following steps:
(I) enhancement pulse and learning algorithm thereof
Unlike the binary form of the pulses, the enhancement pulses can use additional dimensions to carry additional information, such as the number of pulses in a burst of pulses (see fig. 1. C). For convenience of presentation, we abstract the burst of pulses as a pulse coefficient, whereby the information is packed as a single enhancement pulse.
(1) Enhanced neuron model
In standard neuron models, such as the leave integral-and-fire (L IF), the effect of afferent pulses on the neuron membrane potential V (t) is governed by synaptic weights and time constants, which make them unable to handle the dimension of the impulse coefficients in the enhancement pulses.
Figure BDA0002405926110000021
Wherein the content of the first and second substances,
Figure BDA0002405926110000022
is the time to reach the jth pulse of the ith synapse,
Figure BDA0002405926110000023
which is indicative of the corresponding pulse coefficient,
Figure BDA0002405926110000024
representing the time of the jth output pulse of the current neuron. N and wiRepresenting the number of pre-synaptic neurons and the corresponding synaptic weights. θ represents the threshold of the neuron. K (t) is a kernel function defined as:
Figure BDA0002405926110000025
V0is a constant factor used to normalize k (t). Tau ismTime constant, τ, representing the membrane potentialsRepresenting the time constant of the synaptic current.
(2) AugTempotron learning algorithm
Tempotron (tmp) is widely used in various learning tasks due to its simplicity and high efficiency. Following the principles of Tmp, we propose a first learning algorithm, enhanced Tmp (augtmp), to learn and process the enhanced pulse.
In the binary task, each pulse pattern graph belongs to one of two categories (denoted as a and B). The AugTmp learning rule aims to train neurons to fire pulses on the a-mode pattern while keeping B silent. When an error occurs, it will modify the synaptic weights. AugTmp adjusts the weight of neurons by minimizing a loss function, defined as
Figure BDA0002405926110000031
Wherein t ismaxIndicating the point in time at which the neuronal membrane potential reaches its maximum. A gradient descent method is applied to minimize the loss function and then an AugTmp learning algorithm can be derived.
Figure BDA0002405926110000032
Where η is the learning rate.
(3) AugTDP learning algorithm
More recently, multi-pulse learning algorithms have been proposed to train neurons to emit a certain number of pulses. The multi-pulse method exhibits better performance than previous methods. In the present invention, we chose the TDP multi-pulse method to develop a new enhanced multi-pulse learning algorithm (AugTDP) in view of simplicity and efficiency.
The output response of a neuron may be determined by its firing threshold. For example, a lower threshold will generally cause more pulses to be delivered. Thus, the Spike Threshold Surface (STS) function is proposed to describe the relationship between neuron Threshold and number of output pulses. STS defines a series of critical thresholds that change the number of output pulses from k to k-1. Therefore, a critical threshold may be used to adjust the synaptic weights to obtain the desired response. Following the steps in TDP, we propose an enhanced multi-pulse learning rule based on STS. Given a critical threshold value theta*Then it is relative to weight wiDerivative of (2)
Figure BDA0002405926110000033
Can be expressed as
Figure BDA0002405926110000034
Wherein
Figure BDA0002405926110000035
t*Membrane potential represented by θ*The critical time of time. m is t*Total number of previous output pulses. The first term in the above equation can be derived directly, and the last term is zero, so only the second term needs to be solved. According to the procedure of TDP, the above formula can be rewritten as
Figure BDA0002405926110000041
For convenience, we use txTo represent
Figure BDA0002405926110000042
The solving formula of each part in the above formula is as follows.
Figure BDA0002405926110000043
Figure BDA0002405926110000044
Figure BDA0002405926110000045
From this we can get the gradient
Figure BDA0002405926110000046
The direction of the gradient is output by the number n of actual pulsesoAnd the target number ndThe relationship between them.
Figure BDA0002405926110000047
By equation (10), we can adjust the weights of the neurons to obtain the desired number of pulse outputs.
(II) image coding algorithm
To apply enhancement pulses to the image recognition task, we have used three image coding methods to convert the input image into a pulse pattern map.
(1) S1C1 and HMAX process
S1C1 and HMAX are two typical hierarchical time-series coding methods, which integrate information in the receptive field using gaussian difference filter and Gabor filter as weights for coding neurons, respectively.
(2) CNN method
In CNN-based coding methods, the convolutional and pooling layers in the trained CNN are used as coding front-ends and the fully-connected layers are discarded. The network structure of the complete CNN used in the invention is 6C5@28x 28-P2-F256-F10.
(3) Phase transmission time method
In order to generate the launch time of the coding neuron, the invention proposes a phase launch method. Biological experiments have shown that the emission time of the pulse is related to the phase of the membrane potential oscillation. Based on the above, the present invention provides a new phase method to make the coding neuron output sparse and discrete pulse spatiotemporal pattern diagram. If the activation value of a coded neuron is above the firing threshold, the neuron will fire a pulse at some predetermined random point in time within the time window 0 to T100 ms. Notably, in our method, the activation values of the coding neurons are assigned as impulse coefficients. Thus, the input image is encoded as an enhanced pulse pattern map for further learning and classification.
Advantageous effects
The present invention first proposes a new concept of enhancing pulses, in which pulse coefficients are used to represent other information than the pulse firing time. Thereafter, two new learning algorithms are proposed to process the enhancement pulses. The invention provides a plurality of brain-like image recognition methods based on enhanced pulses by combining the current pulse coding method. The experimental result shows that compared with the prior model based on the binary form pulse, the image recognition technology based on the enhanced pulse greatly improves the performance of the SNN on the image recognition task, and highlights the advantages of the enhanced pulse and the learning method. It is noted that the enhanced pulse proposed in the present invention can be easily extended to other pulse-based systems, which is beneficial to further explore more efficient image recognition frameworks and provide new potential research directions for neuromorphic computation.
Drawings
FIG. 1: classical SNN models and graphical illustration of the biological phenomena of pulsed bursts. A, the classical SNN model uses pulses in binary form, i.e. precise pulse firing times, to transmit information. B, burst phenomenon of biological neurons in the nervous system, i.e. several pulses are emitted in a short time. Based on this teaching, the present invention proposes an enhanced pulse model for transmitting information using both pulse emission time and number of emissions (we abstract this number as pulse coefficients).
Fig. 2 is a diagram of the conventional binary form of the pulse space-time and an exemplary diagram of the enhancement pulse proposed by the present invention, with the dot size representing the corresponding pulse coefficient.
Fig. 3 shows a comparison of the present invention with the current state-of-the-art impulse neural network model, with the accuracy in the table based on the MNIST dataset.
Detailed Description
The use of the present invention is described in detail below.
(1) And (5) image time sequence coding.
The input picture is firstly coded into the activation value of the neuron by using three image coding methods of S1C1, HMAX and CNN, and then the pulse emitting time of the coding neuron is generated by using the phase pulse method in the invention, so that the input picture is converted into a pulse space-time diagram for sparseness as the input of the next stage.
(2) Enhanced pulse learning
After the pulse space-time diagrams are obtained through a coding method, the weight of the neuron is adjusted by using an enhanced pulse learning algorithm AugTmp and AugTDP, so that one or a certain number of pulses are emitted to the pulse space-time diagrams of the specified category, and other categories are kept silent.
(I) enhancement pulse and learning algorithm thereof
Unlike the binary form of the pulses, the enhancement pulses can use additional dimensions to carry additional information, such as the number of pulses in a burst of pulses (see fig. 1. C). For convenience of presentation, we abstract the burst of pulses as a pulse coefficient, whereby the information is packed as a single enhancement pulse.
(1) Enhanced neuron model
In standard neuron models, such as the leave integral-and-fire (L IF), the effect of afferent pulses on the neuron membrane potential V (t) is governed by synaptic weights and time constants, which make them unable to handle the dimension of the impulse coefficients in the enhancement pulses.
Figure BDA0002405926110000071
Wherein the content of the first and second substances,
Figure BDA0002405926110000072
is the time to reach the jth pulse of the ith synapse,
Figure BDA0002405926110000073
which is indicative of the corresponding pulse coefficient,
Figure BDA0002405926110000074
representing the time of the jth output pulse of the current neuron. N and wiRepresenting the number of pre-synaptic neurons and the corresponding synaptic weights. θ represents the threshold of the neuron. K (t) is a kernel function defined as:
Figure BDA0002405926110000075
V0is a constant factor used to normalize k (t). Tau ismTime constant, τ, representing the membrane potentialsRepresenting the time constant of the synaptic current.
(2) AugTempotron learning algorithm
Tempotron (tmp) is widely used in various learning tasks due to its simplicity and high efficiency. Following the principles of Tmp, we propose a first learning algorithm, enhanced Tmp (augtmp), to learn and process the enhanced pulse.
In the binary task, each pulse pattern graph belongs to one of two categories (denoted as a and B). The AugTmp learning rule aims to train neurons to fire pulses on the a-mode pattern while keeping B silent. When an error occurs, it will modify the synaptic weights. AugTmp adjusts the weight of neurons by minimizing a loss function, defined as
Figure BDA0002405926110000076
Wherein t ismaxIndicating the point in time at which the neuronal membrane potential reaches its maximum. A gradient descent method is applied to minimize the loss function and then an AugTmp learning algorithm can be derived.
Figure BDA0002405926110000077
Where η is the learning rate.
(3) AugTDP learning algorithm
More recently, multi-pulse learning algorithms have been proposed to train neurons to emit a certain number of pulses. The multi-pulse method exhibits better performance than previous methods. In the present invention, we chose the TDP multi-pulse method to develop a new enhanced multi-pulse learning algorithm (AugTDP) in view of simplicity and efficiency.
The output response of a neuron may be determined by its firing threshold. For example, a lower threshold will generally cause more pulses to be delivered. Therefore, the pulseThe Spike Threshold Surface (STS) function is presented to describe the relationship between neuron Threshold and number of output pulses. STS defines a series of critical thresholds that change the number of output pulses from k to k-1. Therefore, a critical threshold may be used to adjust the synaptic weights to obtain the desired response. Following the steps in TDP, we propose an enhanced multi-pulse learning rule based on STS. Given a critical threshold value theta*Then it is relative to weight wiDerivative of (2)
Figure BDA0002405926110000081
Can be expressed as
Figure BDA0002405926110000082
Wherein
Figure BDA0002405926110000083
t*Membrane potential represented by θ*The critical time of time. m is t*Total number of previous output pulses. The first term in the above equation can be derived directly, and the last term is zero, so only the second term needs to be solved. According to the procedure of TDP, the above formula can be rewritten as
Figure BDA0002405926110000084
For convenience, we use txTo represent
Figure BDA0002405926110000085
The solving formula of each part in the above formula is as follows.
Figure BDA0002405926110000086
Figure BDA0002405926110000087
Figure BDA0002405926110000091
From this we can get the gradient
Figure BDA0002405926110000092
The direction of the gradient is output by the number n of actual pulsesoAnd the target number ndThe relationship between them.
Figure BDA0002405926110000093
By equation (10), we can adjust the weights of the neurons to obtain the desired number of pulse outputs.
(II) image coding algorithm
To apply enhancement pulses to the image recognition task, we have used three image coding methods to convert the input image into a pulse pattern map.
(4) S1C1 and HMAX process
S1C1 and HMAX are two typical hierarchical time-series coding methods, which integrate information in the receptive field using gaussian difference filter and Gabor filter as weights for coding neurons, respectively.
(5) CNN method
In CNN-based coding methods, the convolutional and pooling layers in the trained CNN are used as coding front-ends and the fully-connected layers are discarded. The network structure of the complete CNN used in the invention is 6C5@28x 28-P2-F256-F10.
(6) Phase transmission time method
In order to generate the launch time of the coding neuron, the invention proposes a phase launch method. Biological experiments have shown that the emission time of the pulse is related to the phase of the membrane potential oscillation. Based on the above, the present invention provides a new phase method to make the coding neuron output sparse and discrete pulse spatiotemporal pattern diagram. If the activation value of a coded neuron is above the firing threshold, the neuron will fire a pulse at some predetermined random point in time within the time window 0 to T100 ms. Notably, in our method, the activation values of the coding neurons are assigned as impulse coefficients. Thus, the input image is encoded as an enhanced pulse pattern map for further learning and classification.
Fig. 2 is a diagram of the conventional binary form of the pulse space-time and an exemplary diagram of the enhancement pulse proposed by the present invention, with the dot size representing the corresponding pulse coefficient.
Fig. 3 shows a comparison of the present invention with the current state-of-the-art impulse neural network model, with the accuracy in the table based on the MNIST dataset.

Claims (4)

1. The brain-imitating image identification method based on the enhanced pulse is characterized in that a new concept of the enhanced pulse is provided; afterwards, two new learning algorithms are proposed to process the enhancement pulse; a plurality of brain-imitating image identification methods based on the enhanced pulse are provided by combining the current pulse coding method.
2. The brain-imitated image recognition method based on the enhanced pulse as claimed in claim 1, wherein the enhanced pulse and its learning algorithm: abstracting the pulse bursts into pulse coefficients, whereby the information is packed into a single enhancement pulse;
the enhanced neuron model is as follows:
the L IF model was extended in a standard neuron model to have the ability to read the pulse coefficients:
Figure FDA0002405926100000011
wherein the content of the first and second substances,
Figure FDA0002405926100000012
is the time to reach the jth pulse of the ith synapse,
Figure FDA0002405926100000013
which is indicative of the corresponding pulse coefficient,
Figure FDA0002405926100000014
representing the time of the jth output pulse of the current neuron;
n and wiRepresenting the number of pre-synaptic neurons and corresponding synaptic weights;
θ represents a threshold of the neuron;
k (t) is a kernel function defined as:
Figure FDA0002405926100000015
V0is a constant factor used to normalize k (t);
τmtime constant, τ, representing the membrane potentialsRepresenting the time constant of the synaptic current.
3. The brain-imitated image recognition method based on the enhanced pulse as claimed in claim 1, characterized by comprising two learning algorithms:
(1) AugTempotron learning algorithm
Enhanced tmp (augtmp) to learn and process the enhancement pulse;
in the binary task, each pulse pattern map belongs to one of two categories (denoted as a and B); the AugTmp learning rule aims at training neurons to fire pulses on the a-mode pattern while keeping B silent; when an error occurs, it will modify the synaptic weights; AugTmp adjusts the weights of the neurons by minimizing a loss function, defined as:
Figure FDA0002405926100000021
wherein t ismaxRepresents a point in time at which the neuron membrane potential reaches its maximum value;
applying a gradient descent method to minimize the loss function, then the AugTmp learning algorithm can be derived:
Figure FDA0002405926100000022
wherein η is the learning rate;
(2) AugTDP learning algorithm
The TDP multi-pulse method was chosen to develop a new enhanced multi-pulse learning algorithm (AugTDP).
Following the steps in TDP, we propose an enhanced multipulse learning rule based on STS; given a critical threshold value theta*Then it is relative to weight wiDerivative of (2)
Figure FDA0002405926100000023
Can be expressed as:
Figure FDA0002405926100000024
wherein
Figure FDA0002405926100000025
t*Membrane potential represented by θ*A critical time of time; m is t*The total number of previous output pulses; the first term in the above formula can be obtained by direct derivation, and the last term is zero, so that only the second term needs to be solved;
according to the steps of TDP, the above equation can be rewritten as:
Figure FDA0002405926100000026
using txTo represent
Figure FDA0002405926100000027
The solving formula of each part in the above formula is as follows:
Figure FDA0002405926100000031
Figure FDA0002405926100000032
Figure FDA0002405926100000033
from this we can get the gradient
Figure FDA0002405926100000034
The direction of the gradient is output by the number n of actual pulsesoAnd the target number ndThe relationship between them;
Figure FDA0002405926100000035
by equation (10), we can adjust the weights of the neurons to obtain the desired number of pulse outputs.
4. The method of claim 1, wherein three image coding methods are used to convert the input image into pulse pattern diagram:
(1) S1C1 and HMAX process
S1C1 and HMAX are two typical hierarchical time sequence coding methods, and a Gaussian difference filter and a Gabor filter are respectively adopted as weights of coding neurons to integrate information in a receptive field;
(2) CNN method
In the CNN-based coding method, the convolutional and pooling layers in the trained CNN are used as coding front-ends, and the fully-connected layers are discarded;
the network structure of the complete CNN used in the invention is 6C5@28x 28-P2-F256-F10;
(3) phase transmission time method
A new phase method is provided to make the coding neuron output a sparse and discrete pulse space-time pattern diagram:
if the activation value of a coded neuron is above the firing threshold, the neuron will fire a pulse at some predetermined random point in time within the time window 0 to T100 ms.
CN202010161401.0A 2020-03-10 2020-03-10 Brain-imitating image identification method based on enhanced pulse Pending CN111488908A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010161401.0A CN111488908A (en) 2020-03-10 2020-03-10 Brain-imitating image identification method based on enhanced pulse

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010161401.0A CN111488908A (en) 2020-03-10 2020-03-10 Brain-imitating image identification method based on enhanced pulse

Publications (1)

Publication Number Publication Date
CN111488908A true CN111488908A (en) 2020-08-04

Family

ID=71811710

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010161401.0A Pending CN111488908A (en) 2020-03-10 2020-03-10 Brain-imitating image identification method based on enhanced pulse

Country Status (1)

Country Link
CN (1) CN111488908A (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106845541A (en) * 2017-01-17 2017-06-13 杭州电子科技大学 A kind of image-recognizing method based on biological vision and precision pulse driving neutral net
CN109117884A (en) * 2018-08-16 2019-01-01 电子科技大学 A kind of image-recognizing method based on improvement supervised learning algorithm
CN110210563A (en) * 2019-06-04 2019-09-06 北京大学 The study of pattern pulse data space time information and recognition methods based on Spike cube SNN
CN110751067A (en) * 2019-10-08 2020-02-04 艾特城信息科技有限公司 Dynamic expression recognition method combined with biological form neuron model

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106845541A (en) * 2017-01-17 2017-06-13 杭州电子科技大学 A kind of image-recognizing method based on biological vision and precision pulse driving neutral net
CN109117884A (en) * 2018-08-16 2019-01-01 电子科技大学 A kind of image-recognizing method based on improvement supervised learning algorithm
CN110210563A (en) * 2019-06-04 2019-09-06 北京大学 The study of pattern pulse data space time information and recognition methods based on Spike cube SNN
CN110751067A (en) * 2019-10-08 2020-02-04 艾特城信息科技有限公司 Dynamic expression recognition method combined with biological form neuron model

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
QI XU等: "CSNN: An Augmented Spiking based Framework with Perceptron-Inception", 《RESEARCHGATE》 *
张马路: "Spiking 机器学习算法研究", 《中国博士学位论文全文数据库信息科技辑》 *

Similar Documents

Publication Publication Date Title
CN112633497B (en) Convolutional impulse neural network training method based on re-weighted membrane voltage
CN111858989B (en) Pulse convolution neural network image classification method based on attention mechanism
CN108985252B (en) Improved image classification method of pulse depth neural network
CN109102000A (en) A kind of image-recognizing method extracted based on layered characteristic with multilayer impulsive neural networks
CN112906828A (en) Image classification method based on time domain coding and impulse neural network
CN111310816B (en) Method for recognizing brain-like architecture image based on unsupervised matching tracking coding
CN114266351A (en) Pulse neural network training method and system based on unsupervised learning time coding
Amari Mathematical theory of neural learning
CN113902092A (en) Indirect supervised training method for impulse neural network
CN111488908A (en) Brain-imitating image identification method based on enhanced pulse
Zhang et al. Self-organizing reservoir computing based on spiking-timing dependent plasticity and intrinsic plasticity mechanisms
KR102191346B1 (en) Method for generating spiking neural network based on burst spikes and inference apparatus based on spiking neural network
CN110991602A (en) Event-driven pulse neuron simulation algorithm based on single exponential kernel
Lan et al. Pc-snn: Supervised learning with local hebbian synaptic plasticity based on predictive coding in spiking neural networks
Knoblock et al. Investigation of spiking neural networks for modulation recognition using spike-timing-dependent plasticity
CN113628615B (en) Voice recognition method and device, electronic equipment and storage medium
CN116108891A (en) Lightweight pruning method, device, terminal and computer readable storage medium based on Transformer
CN115602156A (en) Voice recognition method based on multi-synapse connection optical pulse neural network
KR102535635B1 (en) Neuromorphic computing device
CN114239405A (en) Pulse neural network neuron membrane voltage calculation method
Zhang et al. Intrinsic plasticity based inference acceleration for spiking multi-layer perceptron
CN110717590A (en) Efficient multi-pulse algorithm based on single-exponential kernel and neural membrane potential states
Sanchez-Garcia et al. Efficient visual object representation using a biologically plausible spike-latency code and winner-take-all inhibition
Chen et al. Conversion of artificial neural network to spiking neural network for hardware implementation
Zhang et al. A Bionic Spiking Recurrent Neural Network with Sparse Connections and Dale's Principle

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
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20200804

WD01 Invention patent application deemed withdrawn after publication