A hybrid stdp training method and system based on threshold increment cycle count

Abstract

The present application relates to a kind of mixed STDP training method and system based on threshold increment cycle counting, construct including input layer, hidden layer and classification output layer STDP network architecture, and the threshold voltage increment of neuron is counted and controlled to the unsupervised learning process of STDP network architecture, the supervised learning process of combination classification output layer adjusts the weight value corresponding to neuron, by applying threshold increment cycle counting and neuron responsivity screening, effectively constrain weight update, reduce the weight update frequency, to greatly reduce the overall amount of computation of training, help to improve training learning efficiency.The present application method is applicable to handwritten numeral recognition, compared with traditional STDP training algorithm, can reduce 2.5 to 3.5 times amount of computation and improve about 3% classification accuracy.

Description

Technical Field

[0001] This invention relates to the fields of neural network training and artificial intelligence technology, and in particular to a hybrid STDP training method and system based on threshold increment cyclic counting. Background Technology

[0002] Spiking Neural Networks (SNNs), as third-generation neural networks, are characterized by high biomimicry and low power consumption. One common training method for spiking neural networks is based on the pulse time-dependent plasticity (STDP) learning rule. STDP uses a biomimetic model to establish a correlation model between pulse arrival time and weights to change the strength of neuronal synaptic connections. Existing STDP training methods mainly include unsupervised STDP, supervised STDP, and hybrid STDP. Unsupervised STDP performs well in lightweight neural networks but is not suitable for deep networks. Supervised STDP further improves the classification accuracy of the network, but this training method relies on complex mathematical models to improve accuracy. Hybrid STDP combines supervised and unsupervised training methods, first using unsupervised pre-training and then using supervised guidance signals to adjust the model weights, greatly improving the training accuracy and ease of use of the STDP method.

[0003] However, existing hybrid STDP algorithms, due to the prevalence of specific network layers such as convolutional and pooling layers in their network structures, still suffer from narrow applicability, structural limitations, and low learning efficiency. In practical applications, they typically require additional optimization strategies to enhance their adaptability, and their network depth and learning efficiency remain significantly insufficient.

[0004] On the other hand, existing STDP training methods mostly employ lateral inhibition and homeostasis mechanisms to ensure the overall network stability. Lateral inhibition refers to the inhibitory effect of a receiving neuron on other neurons receiving pulses, thus enabling each neuron to effectively learn the corresponding knowledge. A common method for achieving homeostasis is weight scaling, which sets the sum of all synaptic weights connected to postsynaptic neurons as a constant. This not only balances intersynaptic weights, preventing any synaptic weight from becoming infinitely large or zero, but also creates competitive loops between synapses, improving learning efficiency. Another method for achieving homeostasis is adaptive threshold voltage. Specifically, this method increases the threshold voltage whenever the membrane voltage of a LIF neuron exceeds a threshold voltage and fires a pulse. Simultaneously, the sum of increments Based on time constant Exponential decay is employed. The adaptive threshold voltage method can regulate the excitability of neurons, preventing them from continuously firing impulses due to over-excitation.

[0005] However, although existing STDP training methods constrain the weight update frequency by proposing weight scaling and adaptive threshold voltage, they still cannot effectively filter weak response neurons. This inevitably results in a certain number of invalid learning attempts during the learning process, leading to a high weight update frequency, excessive computation, and low learning efficiency. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a hybrid STDP training method and system based on threshold incremental cyclic counting. By applying threshold incremental cyclic counting and neuron responsiveness screening, weight updates are effectively constrained, reducing the frequency of weight updates and thus significantly reducing the overall computational load during training, thereby improving training efficiency. When applied to handwritten digit recognition, compared to the traditional STDP training algorithm, it can reduce computational load by 2.5 to 3.5 times and improve classification accuracy by approximately 3%.

[0007] To achieve the above objectives, the technical solution adopted by the present invention includes:

[0008] A hybrid STDP training method based on threshold increment cyclic counting, characterized by comprising:

[0009] S1. Construct an STDP network architecture, which includes an input layer for receiving pulse signals, a hidden layer for performing nonlinear discrete feature transformation, and a classification output layer.

[0010] S2. Perform an unsupervised learning process on the input layer, hidden layer, and classification output layer, wherein the unsupervised learning process includes membrane voltage updates based on a steady-state mechanism;

[0011] S3. Count the threshold voltage increment of the neuron. When the threshold voltage increment reaches the preset count value, the neuron resends a pulse and resets the threshold voltage increment count.

[0012] S4. Determine whether the current voltage of the neuron is equal to the future voltage. If the current voltage of the neuron is not equal to the future voltage, repeat the unsupervised learning process.

[0013] S5. When it is determined that the current voltage of the neuron is equal to the future voltage, the unsupervised learning process ends and the first training result is obtained.

[0014] Furthermore, the hybrid STDP training method also includes:

[0015] S6. Perform a supervised learning process on the classification output layer to obtain a second training result. The supervised learning process includes membrane voltage updates based on the guidance signal.

[0016] S7. Use the second training result to correct the first training result and adjust the weight values ​​of the corresponding neurons.

[0017] S8. Determine whether a neuron is a weak response neuron based on the adjusted weight value, and remove the identified weak response neurons.

[0018] Furthermore, the STDP network architecture includes a network architecture based on LIF neurons, as shown in Equation 1:

[0019] Formula 1

[0020] In Equation 1, For leakage constant, It is the neuronal membrane voltage. It is the threshold voltage. It's a voltage reset. For film capacitors, For membrane resistance, This represents the membrane voltage at the current moment. This represents the membrane voltage at the previous moment. It is the membrane voltage leakage term. Represents the pulse emission function. Neuron and neurons Connected synaptic weights.

[0021] Furthermore, the membrane voltage update based on the steady-state mechanism is shown in Equation 2:

[0022] Formula 2

[0023] In Equation 2, Represents presynaptic neurons The membrane potential on the surface, Represents postsynaptic neurons The threshold voltage on, It is a training time window. It is the total training time. This represents the updated membrane potential value in steady state. This is the updated value of the postsynaptic membrane potential, in relation to... Neurons connected to each other Above, current voltage and future voltage They will converge.

[0024] Furthermore, the membrane voltage update based on the guidance signal is shown in Equation 3:

[0025] Formula 3

[0026] In Equation 3, As a guidance signal, The current voltage, and These are the first and second impact factors, respectively, used to control parameter updates.

[0027] Furthermore, the supervised learning process also includes screening based on neuron responsiveness, as shown in Equation 4:

[0028] Formula 4

[0029] In Equation 4, For the test image set, For training image sets, For all tags The set of neurons that respond. For neurons For tags The number of pulses, For the sudden trigger after the weight reduction , Weights on connected neurons This is the weighting factor.

[0030] Further, step S7 includes:

[0031] Use formula 5 to update synaptic weights.

[0032] Formula 5

[0033] In Equation 5, It is a postsynaptic neuron The membrane potential on the surface, It is a presynaptic neuron The derivative of the upper membrane potential, It is the third most influential factor.

[0034] This invention also relates to a hybrid STDP training system based on threshold increment cyclic counting, characterized in that it includes:

[0035] The network setup module is used to build the STDP network architecture;

[0036] The unsupervised learning module is used to perform unsupervised learning processes on the input layer, hidden layer, and classification output layer to obtain the first training result.

[0037] The supervised learning module is used to perform a supervised learning process on the classification output layer to obtain a second training result;

[0038] The weight update module is used to correct the first training result using the second training result and adjust the weight values ​​corresponding to the neurons.

[0039] The neuron identification module is used to determine whether a neuron is a weak-response neuron based on the adjusted weight value, and to remove the identified weak-response neurons.

[0040] The present invention also relates to a computer-readable storage medium, characterized in that the storage medium stores a computer program, which, when executed by a processor, implements the above-described method.

[0041] The present invention also relates to an electronic device, characterized in that it includes a processor and a memory;

[0042] The memory is used to store the threshold voltage increment, the first training result, the second training result, and the weight value;

[0043] The processor is configured to execute the above method by invoking the threshold voltage increment, the first training result, the second training result, and the weight value.

[0044] The present invention also relates to a computer program product, including a computer program and / or instructions, characterized in that the computer program and / or instructions, when executed by a processor, implement the steps of the above-described method.

[0045] The beneficial effects of this invention are as follows:

[0046] The hybrid STDP training method and system based on threshold incremental cyclic counting described in this invention addresses the shortcomings of existing STDP training methods, such as the inability to eliminate weak-response neurons, leading to high weight update frequency and low learning efficiency. By applying threshold incremental cyclic counting and neuron responsiveness screening, it effectively constrains weight updates, reduces the weight update frequency, and eliminates weak-response neurons, thereby significantly reducing the overall computational load of training and improving training efficiency. Applicable to handwritten digit recognition, compared to traditional STDP training algorithms, it can reduce computational load by 2.5 to 3.5 times and improve classification accuracy by approximately 3%. It is a universal, scalable, and efficient fully connected neural network framework. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the first embodiment of the hybrid STDP training method based on threshold increment cyclic counting of the present invention.

[0048] Figure 2 This is a schematic diagram of the second embodiment of the hybrid STDP training method based on threshold increment cyclic counting of the present invention.

[0049] Figure 3 This is a schematic diagram of the hybrid STDP training system based on threshold increment cyclic counting according to the present invention. Detailed Implementation

[0050] To better understand the content of this invention, a detailed description will be provided in conjunction with the accompanying drawings and embodiments.

[0051] The first aspect of this invention relates to a process flow as follows: Figure 1 as well as Figure 2 The hybrid STDP training method based on threshold increment cyclic counting, as shown, includes:

[0052] S1. Construct an STDP network architecture, which includes an input layer for receiving pulse signals, a hidden layer for performing nonlinear discrete feature transformation, and a classification output layer.

[0053] Specifically, the STDP network architecture consists of three layers: the first layer is the input layer that receives pulse signals, the middle layer is the hidden layer for nonlinear discrete feature transformation, and the last layer is the classification output layer with supervised learning signals. Preferably, the middle hidden layer network can be extended and deepened to extract more features.

[0054] Preferably, the STDP network architecture uses LIF neurons to construct the network. LIF neurons are a commonly used spiking neural network neuron model. Their operating rules are as follows: the neuron receives a presynaptic pulse signal input; with the arrival of the pulse signal, the neuron's membrane voltage accumulates; after the accumulated voltage reaches a threshold, the LIF neuron fires a pulse, and the membrane voltage becomes a reset voltage. The LIF neuron is shown in Equation 1:

[0055] Formula 1

[0056] In Equation 1, For leakage constant, It is the neuronal membrane voltage. It is the threshold voltage. It's a voltage reset. For film capacitors, For membrane resistance, This represents the membrane voltage at the current moment. This represents the membrane voltage at the previous moment. It is the membrane voltage leakage term. Represents the pulse emission function. Neuron and neurons Connected synaptic weights.

[0057] Preferably, for applications involving handwritten digit recognition, the pulse signal input can be W. A grayscale image of a handwritten digit H is generated. Each pixel in the image is Poisson encoded using its pixel intensity as the pulse emission rate, forming a pulse sequence with different emission frequencies, which is then sent one-to-one to a quantity of K=W. H neurons.

[0058] S2. Perform an unsupervised learning process on the input layer, hidden layer, and classification output layer, wherein the unsupervised learning process includes membrane voltage updates based on a steady-state mechanism.

[0059] The unsupervised learning process described above is based on steady-state membrane voltage updates. That is, during unsupervised learning, the input and output voltages of neurons are kept in balance, and as the neural network training progresses, the voltage updates in accordance with the steady-state mechanism. Neurons connected to each other Above, current voltage and future voltage They will converge until they are equal at the end of training. The membrane voltage update based on the steady-state mechanism is shown in Equation 2:

[0060] Formula 2

[0061] In Equation 2, Represents presynaptic neurons The membrane potential on the surface, Represents postsynaptic neurons The threshold voltage on, It is a training time window. It is the total training time. This represents the updated membrane potential value in steady state. This is the updated value of the postsynaptic membrane potential, in relation to... Neurons connected to each other Above, current voltage and future voltage They will converge.

[0062] S3. Count the threshold voltage increment of the neuron. When the threshold voltage increment reaches the preset count value, the neuron resends a pulse and resets the threshold voltage increment count.

[0063] Specifically, LIF neurons continuously accumulate membrane voltage based on the arrival of presynaptic pulses. When the membrane voltage exceeds the threshold voltage, a pulse signal is emitted postsynaptically, resetting the voltage to the resting state. At this point, according to the adaptive membrane voltage threshold change rule of the homeostasis mechanism, the neuron's threshold voltage increases. To prevent neurons from continuously firing pulses due to over-excitation, the threshold voltage increment of the neuron is counted. A pulse is only fired when the threshold voltage increment reaches the count value, after which the counting restarts. Unlike existing adaptive threshold voltage methods, the threshold increment cyclic counting method... Perform cyclic counting, with the main component being the threshold increment. The adaptive threshold voltage method modifies the threshold voltage, with the threshold voltage as the main component. .

[0064] S4. Determine whether the current voltage of the neuron is equal to the future voltage. If the current voltage of the neuron is not equal to the future voltage, repeat the unsupervised learning process.

[0065] S5. When it is determined that the current voltage of the neuron is equal to the future voltage, the unsupervised learning process ends and the first training result is obtained.

[0066] S6. Perform a supervised learning process on the classification output layer to obtain a second training result. The supervised learning process includes membrane voltage updates based on the guidance signal.

[0067] Specifically, to facilitate control over supervised learning efficiency and membrane voltage adjustment range, an influencing factor is added. and The control parameters are updated, and the membrane voltage update based on the guidance signal is shown in Equation 3:

[0068] Formula 3

[0069] In Equation 3, As a guidance signal, The current voltage, and These are the first and second impact factors, respectively, used to control parameter updates.

[0070] Preferably, the supervised learning process further includes screening based on neuron responsiveness. Neuron responsiveness screening requires calculating the weight update formula based on the number of pulses fired by neurons in the test results. Neurons that fired pulses during training but not during testing are screened based on the test results, and the weights of these neurons are reduced to eliminate weakly responsive neurons and improve the responsiveness of the remaining neurons to the classification label. The screening based on neuron responsiveness is shown in Equation 4:

[0071] Formula 4

[0072] In Equation 4, For the test image set, For training image sets, For all tags The set of neurons that respond. For neurons For tags The number of pulses, For the sudden trigger after the weight reduction , Weights on connected neurons This is the weighting factor.

[0073] S7. Use the second training result to correct the first training result, and adjust the weight values ​​of the corresponding neurons. Specifically, this includes updating the synaptic weights using Equation 5.

[0074] Formula 5

[0075] In Equation 5, It is a postsynaptic neuron The membrane potential on the surface, It is a presynaptic neuron The derivative of the upper membrane potential, It is the third most influential factor.

[0076] S8. Determine whether a neuron is a weak response neuron based on the adjusted weight value, and remove the identified weak response neurons.

[0077] When using the above method for STDP training, specific count values ​​can be set for corresponding control. For example, when the count value is set to 3, each time the threshold voltage increases... After three firings, the neuron fires a pulse and restarts the threshold increment count. If a neuron's pulse count was 10 during training but only 1 during testing, the weights on that neuron will be reduced to their original values. Furthermore, with repeated weight updates, the weights on this neuron gradually approach 0, thereby eliminating weak-response neurons.

[0078] Another aspect of the present invention relates to a hybrid STDP training system based on threshold increment cyclic counting, the structure of which is as follows: Figure 3 As shown, it includes:

[0079] The network setup module is used to build the STDP network architecture;

[0080] The unsupervised learning module is used to perform unsupervised learning processes on the input layer, hidden layer, and classification output layer to obtain the first training result.

[0081] The supervised learning module is used to perform a supervised learning process on the classification output layer to obtain a second training result;

[0082] The weight update module is used to correct the first training result using the second training result and adjust the weight values ​​corresponding to the neurons.

[0083] The neuron identification module is used to determine whether a neuron is a weak-response neuron based on the adjusted weight value, and to remove the identified weak-response neurons.

[0084] By using this system, the aforementioned computational processing methods can be executed and the corresponding technical effects can be achieved.

[0085] Embodiments of the present invention also provide a computer-readable storage medium capable of implementing all the steps of the methods in the above embodiments, wherein the computer-readable storage medium stores a computer program that, when executed by a processor, implements all the steps of the methods in the above embodiments.

[0086] Embodiments of the present invention also provide an electronic device for performing the above-described method. As an implementation device for the method, the electronic device includes at least a processor and a memory. In particular, the memory stores data and related computer programs required for performing the method, such as threshold voltage increments, first training results, second training results, and weight values. The processor calls the data and programs in the memory to execute all steps of the method and obtain the corresponding technical effects.

[0087] Preferably, the electronic device may include a bus architecture, which may include any number of interconnected buses and bridges. The bus will include various circuits linked together by one or more processors and memories. The bus may also link together various other circuits such as peripherals, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides an interface between the bus and the receiver and transmitter. The receiver and transmitter may be the same element, i.e., a transceiver, providing a unit for communicating with various other systems over a transmission medium. The processor is responsible for managing the bus and general processing, while the memory may be used to store data used by the processor during operation.

[0088] Additionally, the electronic device may further include a communication module, an input unit, an audio processor, a display, a power supply, and other components. The processor (or controller, operating control) used may include a microprocessor or other processor device and / or logic device, which receives input and controls the operation of various components of the electronic device. The memory may be one or more of a buffer, flash memory, hard drive, removable media, volatile memory, non-volatile memory, or other suitable devices, which can store the aforementioned data information, and may also store programs for executing the information. The processor can execute the program stored in the memory to achieve information storage or processing, etc. The input unit is used to provide input to the processor, for example, it can be a button or touch input device. The power supply is used to provide power to the electronic device. The display is used to display images and text, for example, it can be an LCD display. The communication module is a transmitter / receiver that transmits and receives signals via an antenna. The communication module (transmitter / receiver) is coupled to the processor to provide input signals and receive output signals, which can be the same as in conventional mobile communication terminals. Based on different communication technologies, multiple communication modules can be incorporated into the same electronic device, such as cellular network modules, Bluetooth modules, and / or wireless LAN modules. The communication module (transmitter / receiver) is also coupled to a speaker and microphone via an audio processor to provide audio output through the speaker and receive audio input from the microphone, thereby enabling typical telecommunications functions. The audio processor can include any suitable buffer, decoder, amplifier, etc. Furthermore, the audio processor is coupled to a central processing unit, enabling on-device recording via the microphone and on-device playback of stored sound via the speaker.

[0089] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0090] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0091] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0092] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the functions specified in one or more boxes. Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0093] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A hybrid STDP training method based on threshold increment cyclic counting, characterized in that, Used for handwritten digit recognition, including: S1. Construct an STDP network architecture, which includes an input layer for receiving pulse signals, a hidden layer for performing nonlinear discrete feature transformation, and a classification output layer. The input of the pulse signal is W. H is a grayscale image of a handwritten digit. Each pixel of the grayscale image is Poisson encoded using pixel intensity as the pulse emission rate, forming pulse sequences with different emission frequencies that are sent one-to-one to a quantity of K=W. H neurons; S2. An unsupervised learning process is performed on the input layer, hidden layer, and classification output layer. This unsupervised learning process includes membrane voltage updates based on a steady-state mechanism. During the unsupervised learning process, the input and output voltages of the neurons are kept in balance. Neurons connected to each other Above, current voltage and future voltage They will approach consistency until they are equal at the end of training. The membrane voltage update based on the steady-state mechanism is shown in Equation 2: Formula 2 In Equation 2, Represents presynaptic neurons The membrane potential on the surface, Represents postsynaptic neurons The threshold voltage on, It is a training time window. It is the total training time. This represents the updated membrane potential value in steady state. This is the updated value of the postsynaptic membrane potential; S3. Count the threshold voltage increment of the neuron. When the threshold voltage increment reaches the preset count value, the neuron resends a pulse and resets the threshold voltage increment count. S4. Determine whether the current voltage of the neuron is equal to the future voltage. If the current voltage of the neuron is not equal to the future voltage, repeat the unsupervised learning process. S5. When it is determined that the current voltage of the neuron is equal to the future voltage, the unsupervised learning process ends and the first training result is obtained.

2. The method as described in claim 1, characterized in that, The hybrid STDP training method also includes: S6. Perform a supervised learning process on the classification output layer to obtain a second training result. The supervised learning process includes membrane voltage updates based on the guidance signal. S7. Use the second training result to correct the first training result and adjust the weight values ​​of the corresponding neurons. S8. Determine whether a neuron is a weak response neuron based on the adjusted weight value, and remove the identified weak response neurons.

3. The method as described in claim 2, characterized in that, The STDP network architecture includes a LIF neuron-based network architecture, as shown in Equation 1: Formula 1 In Equation 1, For leakage constant, It is the neuronal membrane voltage. It is the threshold voltage. It's a voltage reset. For film capacitors, For membrane resistance, This represents the membrane voltage at the current moment. This represents the membrane voltage at the previous moment. It is the membrane voltage leakage term. Represents the pulse emission function. Neuron and neurons Connected synaptic weights.

4. The method as described in claim 3, characterized in that, The membrane voltage update based on the guidance signal is shown in Equation 3: Formula 3 In Equation 3, As a guidance signal, The current voltage, and These are the first and second impact factors, used to control parameter updates.

5. The method as described in claim 4, characterized in that, The supervised learning process further includes screening based on neuron responsiveness, as shown in Equation 4: Formula 4 In Equation 4, For the test image set, For training image set, For all tags The set of neurons that make a response. For neurons For tags The number of pulses, For the sudden trigger after the weight reduction , Weights on connected neurons This is the weighting factor.

6. The method as described in claim 3, characterized in that, Step S7 includes: Use formula 5 to update synaptic weights. Formula 5 In Equation 5, It is a postsynaptic neuron The membrane potential on the surface, It is a presynaptic neuron The derivative of the upper membrane potential, It is the third most influential factor.

7. A hybrid STDP training system based on threshold increment cyclic counting, characterized in that, For implementing the method as described in any one of claims 1-6, for handwritten digit recognition, comprising: The network setup module is used to build the STDP network architecture; The unsupervised learning module is used to perform unsupervised learning processes on the input layer, hidden layer, and classification output layer to obtain the first training result. The supervised learning module is used to perform a supervised learning process on the classification output layer to obtain a second training result; The weight update module is used to correct the first training result using the second training result and adjust the weight values ​​corresponding to the neurons. The neuron identification module is used to determine whether a neuron is a weak-response neuron based on the adjusted weight value, and to remove the identified weak-response neurons.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method of any one of claims 1 to 6.

9. An electronic device, characterized in that, Including processor and memory; The memory is used to store the threshold voltage increment, the first training result, the second training result, and the weight value; The processor is configured to execute the method of any one of claims 1 to 6 by invoking the threshold voltage increment, the first training result, the second training result, and the weight value.

10. A computer program product comprising a computer program and / or instructions, characterized in that, When the computer program and / or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 6.

Images