Time adaptive spiking neural network accelerator optimization method, medium and equipment

Abstract

The invention provides a time adaptive spiking neural network accelerator optimization method, a medium and equipment, and the method comprises the steps: calculating the pulse activity degree of a spiking neural network model in a spiking neural network accelerator, so as to measure the contribution degree of each layer in the spiking neural network model to the spiking neural network model; determining a loss function of the pulse neural network model based on the pulse activeness; performing iterative training on the spiking neural network model based on a data set and a loss function to determine a corresponding time step length for each layer of the spiking neural network model; and the spiking neural network accelerator executes an application program task based on the trained spiking neural network model. According to the method, the pulse activeness of each layer of the network is calculated, so that the criticality of each layer in the SNN network is effectively evaluated, the contribution of each layer in the overall network performance is helped to be determined, the optimization of the time step length is guided, and complete information is transmitted among the SNN layers with different time step lengths.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and particularly to the field of spiking neural network technology. Background Art

[0002] Spiking Neural Networks (SNNs) are a type of neural network that mimics the working principle of the biological nervous system. They use spike trains with temporal coding for information processing. This unique information processing method gives SNNs unique advantages in terms of energy efficiency, data processing, and biological realism. Compared with traditional artificial neural networks (ANNs), SNNs simulate the spike firing behavior of biological neurons and perform well in processing time series data, spatio-temporal analysis, and complex pattern recognition tasks. In neuromorphic hardware, SNNs achieve low-power computing by using simple addition operations instead of multiplication operations, which makes them particularly suitable for fields such as speech recognition, machine vision, and robot control.

[0003] Spiking Neural Networks (SNNs) can be divided into two categories according to different information representation methods: SNN-R and SNN-T. In SNN-R, information is represented by the spike count in each time step. The more spikes there are, the more information is expressed. In other words, the information intensity of each layer is proportional to the number of its spikes. On the contrary, in SNN-T, information is represented by the time when the first spike appears in each time step. In this type of network, the earlier the spike appears, the more information it represents. These two unique information representation methods of SNNs endow them with different advantages and characteristics when processing different types of data.

[0004] Although spiking neural networks perform well in some aspects, they also face some key performance bottlenecks:

[0005] 1. Fixed time step: During the inference process, existing SNN methods usually cannot dynamically adjust the time dimension, restricting the adaptability of the time window between different layers of the network. Among them, the size of the time window for neurons to receive input spikes is equal to the number of time steps.

[0006] 2. Inflexible information transmission: Currently, SNNs lack an efficient method to implement information conversion between different time steps between layers, which may lead to information loss or computational redundancy, affecting the model performance and accuracy.

[0007] 3. Application limitations: Due to the above limitations, the performance of SNNs in practical applications may not meet expectations, especially in tasks that require low power consumption and low latency. Summary of the Invention

[0008] In view of the above-mentioned disadvantages of the prior art, the purpose of the present invention is to provide a time-adaptive spiking neural network accelerator optimization method, medium and device for improving the computing efficiency and energy efficiency of the neural network accelerator.

[0009] To achieve the above object and other related objects, the present invention provides a time-adaptive spiking neural network accelerator optimization method, the method comprising: calculating the pulse activity of the spiking neural network model in the spiking neural network accelerator to measure the contribution degree of each layer in the spiking neural network model to the spiking neural network model; determining the loss function of the spiking neural network model based on the pulse activity; performing iterative training on the spiking neural network model based on the data set and the loss function to determine corresponding time steps for each layer of the spiking neural network model; and the spiking neural network accelerator executing an application program task based on the trained spiking neural network model.

[0010] In an embodiment of the present invention, the pulse activity is the ratio of the internal excitation of the current neuron to the length of its time step.

[0011] In an embodiment of the present invention, it further comprises: determining the transmitted pulse sequence for each layer based on the potential threshold of the excitation pulse and the excitation potential of the pulse.

[0012] In an embodiment of the present invention, for the SNN-T type of spiking neural network model, the information transmitted to the next layer is a sequence with a single pulse at the time point where is the time for generating the pulse for compensation update in the SNN, x′ T =x T +V th / V fire where x T is the time for generating the pulse in the SNN-T, V th is the potential threshold of the excitation pulse in the SNN, V fire is the average excitation potential of all pulses in the SNN-T; T l is the time step of the current layer, and T l+1 is the time step of the next layer.

[0013] In an embodiment of the present invention, for the SNN-R type of spiking neural network model, the information transmitted to the next layer is a sequence with pulses at the previous time points, where x′ R is the time for generating the pulse for compensation update in the SNN, x′ R =x R +1-V th / V′ fire where x Ris the pulse count in SNN-R, V th is the potential threshold of the firing pulse in SNN, V′ fire is the firing potential of the only pulse in SNN-R.

[0014] In an embodiment of the present invention, the loss function L of the pulse neural network model is: L = L CE + τl r ; where Act(l) represents the pulse activity of the l-th layer, reflecting the impact of this layer on the network accuracy, Para(l) represents the total number of parameters of the l-th layer, reflecting the computational load of this layer, and L CE is the cross-entropy loss function for the original SNN training, τ is the penalty factor for balancing the cross-entropy loss and the regularization term loss, and l r is the regularization term loss function.

[0015] In an embodiment of the present invention, determining the corresponding time step for each layer of the pulse neural network model includes: in each training cycle, iteratively processing each layer and its respective time steps, and updating the membrane potential of the neuron based on the input pulse sequence and the cumulative input current; updating the loss function and the network weights; updating the time steps of each layer.

[0016] In an embodiment of the present invention, the training process starts with the initialization of the time step for each layer, which is dynamically set based on the pulse activity of each layer.

[0017] To achieve the above and other related purposes, the present invention also provides a computer storage medium storing program instructions, and when the program instructions are executed, the steps of the above-mentioned time-adaptive pulse neural network accelerator optimization method are implemented.

[0018] To achieve the above and other related purposes, the present invention also provides an electronic device including a memory for storing a computer program; and a processor for running the computer program to implement the steps of the above-mentioned time-adaptive pulse neural network accelerator optimization method.

[0019] As described above, the time-adaptive pulse neural network accelerator optimization method, medium, and device of the present invention have the following beneficial effects:

[0020] By calculating the pulse activity of each layer of the network, the present invention effectively evaluates the criticality of each layer in the SNN network, helps determine the contribution of each layer to the overall network performance, thereby guiding the optimization of the time step, and realizing the transfer of complete information between SNN layers with different time steps, effectively improving the computational efficiency and energy efficiency of the neural network accelerator. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0022] Figure 1 It shows a schematic diagram of the overall process of the time - adaptive spiking neural network accelerator optimization method in an embodiment of the present application;

[0023] Figure 2 It shows a schematic diagram of the principle of the dynamic pulse train conversion method in the time - adaptive spiking neural network accelerator optimization method in an embodiment of the present application;

[0024] Figure 3 It shows a schematic diagram of the structure of an electronic device in an embodiment of the present application.

[0025] Element number description

[0026] 100 Electronic device

[0027] 101 Memory

[0028] 102 Processor

[0029] 103 Display

[0030] Steps S110 - S400 Detailed implementation manners

[0031] The following uses specific specific examples to illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0032] The purpose of this embodiment is to provide a time - adaptive spiking neural network accelerator optimization method, medium and device, which are used to improve the computing efficiency and energy efficiency of the neural network accelerator.

[0033] The following will elaborate in detail on the principles and implementation manners of the time - adaptive spiking neural network accelerator optimization method, medium and device of the present invention, so that those skilled in the art can understand the time - adaptive spiking neural network accelerator optimization method, medium and device of the present invention without creative efforts.

[0034] This embodiment provides an optimization method for a time - adaptive spiking neural network accelerator. Specifically, as Figure 1 shown, the time - adaptive spiking neural network accelerator optimization method in this embodiment includes:

[0035] Step S100, calculate the spike activity of the spiking neural network model in the spiking neural network accelerator to measure the contribution degree of each layer in the spiking neural network model to the spiking neural network model;

[0036] Step S200, determine the loss function of the spiking neural network model based on the spike activity;

[0037] Step S300, perform iterative training on the spiking neural network model based on the data set and the loss function to determine corresponding time steps for each layer of the spiking neural network model;

[0038] Step S400, the spiking neural network accelerator executes the application program task based on the trained spiking neural network model.

[0039] An optimization method for a time - adaptive spiking neural network accelerator in this embodiment effectively evaluates the key importance of each layer in the SNN network by calculating the spike activity of each layer of the network, helps to determine the contribution of each layer to the overall network performance, thereby guiding the optimization of time steps, and effectively improving the computing efficiency and energy efficiency of the neural network accelerator.

[0040] The above steps S100 to S400 of the time - adaptive spiking neural network accelerator optimization method in this embodiment are described in detail below.

[0041] Step S100, calculate the spike activity of the spiking neural network model in the spiking neural network accelerator to measure the contribution degree of each layer in the spiking neural network model to the spiking neural network model.

[0042] This embodiment adopts the dynamic spike train conversion method (TEAS) to enable SNNs to adapt to different time steps during operation. By adjusting the information characteristics in the time dimension of the SNN, without significantly reducing the model accuracy, redundant calculations are reduced, and the computing efficiency and energy efficiency of the SNN model are improved, thereby improving the computing efficiency and energy efficiency of the spiking neural network accelerator. As Figure 2 shown, the core idea of TEAS is to configure independent time steps for each layer of the SNN based on the contribution of each layer to the overall network performance, so as to fully exert the potential of the SNN, generate a dynamic strategy through the time conversion module, and dynamically adapt to time information during the inference process.

[0043] In this embodiment, for spiking neural networks (SNNs), different layers contribute differently to the model accuracy. By reducing the time steps of the layers that contribute less to the model accuracy while trying to retain the time steps of the layers that contribute more to the model accuracy, the overall accuracy of the model can be maintained while reducing redundant calculations. To measure the impact of each layer on the model accuracy, this embodiment introduces an index called "spike activity (Act)". The layers with higher spike activity contribute more significantly to the model accuracy and exhibit higher sensitivity.

[0044] Specifically, in this embodiment, the spike activity is the ratio of the internal excitation of the current neuron to the length of its time step. That is, the spike activity (Act) is defined as the ratio of the internal excitation (fired spikes) of the current neuron to the length of its time step (time step), i.e.:

[0045] Act(l) = fired spike / time step.

[0046] This embodiment effectively evaluates the criticality of each layer in the SNN network by calculating the spike activity of each layer of the network, helps to determine the contribution of each layer to the overall network performance, and thus realizes the optimization of the time steps of the spiking neural network accelerator.

[0047] In spiking neural networks (SNNs), when the time steps between layers are different, the direct transmission of information becomes complex. Specifically, if the time step of the current layer is T l , and the time step of the next layer is T l+1 , when T l < T l+1 , in the time period from T l+1 - T l , the next layer cannot receive information, resulting in these time steps becoming redundant; and when T l > T l+1 , in the time period from T l - T l+1 , the information of the current layer will be lost and cannot be captured by the next layer. To solve this problem, this embodiment designs an adaptive scaling framework to adaptively change the size required for each layer to transmit the spike train according to the importance of the information.

[0048] Specifically, in this embodiment, it further includes: determining the spike train transmitted by each layer based on the potential threshold of the fired spike and the firing potential of the spike. This embodiment designs a linear mapping to achieve the transmission of complete information between SNN layers with different time steps, as follows:

[0049] 1) In this embodiment, for the spiking neural network model of the SNN-T type, the information passed to the next layer is a sequence with a single pulse at the time point , where x' T is the time for generating a pulse for compensation update in the SNN, aiming to reduce information loss when the time window of the next layer is larger than that of the current layer. x' T = x T + V th / V fire , x T is the time for generating a pulse in the SNN-T, V th is the potential threshold for firing a pulse in the SNN, and V fire is the average firing potential of all pulses in the SNN-T; T l is the time step of the current layer, and T l+1 is the time step of the next layer.

[0050] 2) In this embodiment, for the spiking neural network model of the SNN-R type, the information passed to the next layer is a sequence with pulses at the previous time points, where x' R is the time for generating a pulse for compensation update in the SNN, aiming to reduce information loss when the time window of the next layer is larger than that of the current layer. x' R = x R + 1 - V th / V' fire , x R is the pulse count in the SNN-R, V th is the potential threshold for firing a pulse in the SNN, and V' fire is the firing potential of the only pulse in the SNN-R.

[0051] Through the above method, this embodiment can effectively transfer information between layers with different time steps, and at the same time consider the firing potential of neurons when generating pulses, which reflects the information intensity of neurons.

[0052] Step S200, determine the loss function of the spiking neural network model based on the pulse activity.

[0053] In this embodiment, the loss function L of the spiking neural network model is:

[0054] L = L CE + τl r ;

[0055]

[0056] Among them, Act(l) represents the pulse activity of the l-th layer, reflecting the impact of this layer on the network accuracy, and Para(l) represents the total number of parameters of the l-th layer, reflecting the computational load of this layer, L CE is the cross-entropy loss function for the original SNN training, τ is the penalty factor that balances the cross-entropy loss and the regularization term loss, and l r is the regularization term loss function.

[0057] To enable the spiking neural network (SNN) to dynamically adjust and reduce the time step of non-critical layers during training, this embodiment introduces a special regularization loss term, that is, the regularization term loss function l r The purpose of this regularization loss term is to optimize the network structure and ensure that computational resources are concentrated on the critical layers that contribute more to the model accuracy.

[0058] It can be seen that this embodiment designs a regularization term to encourage reducing the time step of non-critical layers during training, while retaining the time step of critical layers, as well as an SNN training method for dynamically updating the time step of each layer.

[0059] Step S300, iteratively train the spiking neural network model based on the dataset and the loss function to determine corresponding time steps for each layer of the spiking neural network model.

[0060] In this embodiment, determining the corresponding time step for each layer of the spiking neural network model includes: in each training cycle, iteratively process each layer and its respective time steps, and update the membrane potential of neurons based on the input pulse sequence and the cumulative input current; update the loss function and the network weights; update the time steps of each layer.

[0061] In this embodiment, the training process starts with the initialization of the time step for each layer, which is dynamically set based on the pulse activity of each layer.

[0062] That is, in this embodiment, the training process starts with the initialization of the time step for each layer, which is dynamically set based on the pulse activity of each layer. Subsequently, through multiple training cycles, the time step of each layer undergoes continuous evaluation and adjustment. In each training cycle, iteratively process each layer and its respective time steps, and update the membrane potential of neurons by integrating the input pulse train and the cumulative input current. Then, calculate the network loss through forward propagation and calculate the gradient through backward propagation to update the network weights. Then, sort the pulse activities of each layer of the network and adjust the time steps of the top K layers before sorting.

[0063] Step S400, the spiking neural network accelerator executes the application program task based on the trained spiking neural network model.

[0064] When this embodiment is specifically running, for example, based on the Pytorch framework, using CIFAR-10, CIFAR-100, and ImageNet datasets, adopting convolutional SNNs (i.e., VGG-16 and ResNet network structures), TEAS is evaluated on a 4-way NVIDIA Tesla V100 server. For example, for ResNet-20 on CIFAR-10, compared with the original SNN (time step = 2500), TEAS achieves a 61.7-fold reduction in the number of time steps with almost no accuracy loss. For CIFAR-100, the TEAS algorithm has only a 0.02% accuracy loss compared with the original SNN, and compared with the state-of-the-art time compression method

[18] , it achieves a significant 6.3% accuracy improvement while reducing the number of time steps by 40%. For the more complex task of ImageNet, the layer-wise adaptive time step makes our TEAS more robust than the small-window SNN

[19] in maintaining the accuracy of the SNN. Specifically, compared with the method in

[19] , TEAS achieves a 4.9% accuracy improvement while reducing the number of time steps by 67%. This is because the existing efforts to reduce time information are unified across all layers, and this method is too aggressive to achieve good accuracy.

[0065] For the conversion of ANNs with the VGG-16 network structure to SNNs, this embodiment reduces the average time step to 40.5 for CIFAR-10, 71.3 for CIFAR-100, and 77.3 for ImageNet respectively, with an average accuracy loss of only 0.03%, and the maximum accuracy loss is 0.05%. The layer-wise time step results of the SNNs with the VGG-16 / ResNet-20 network structure under our TEAS are shown in Figure 3 Figure. It can be observed that in the SNNs of VGG-16 and ResNet-20, the maximum time steps are 108 and 72 respectively. This embodiment can easily find a better balance between accuracy and the size of the time dimension for model inference accuracy.

[0066] Energy efficiency comparison with traditional ANNs and SNNs with fixed time steps: For VGG-16, compared with ANNs and SNNs with the same time dimension, the SNN adopting TEAS has an energy efficiency improvement of about 11.5 times. This efficiency improvement is mainly attributed to TEAS, which can optimally allocate the time dimension for each layer of the SNN, avoiding resource waste. Mainly because addition operations consume significantly less energy than multiplication operations, the energy efficiency of SNNs is better than that of ANNs, and the SNN adopting TEAS further reduces the addition operations in non-critical layers.

[0067] However, the energy consumption of ANNs depends on the number of floating-point operations (FLOPs) performed, while the energy consumption of SNNs is mainly determined by the number of emitted pulses. Since SNNs usually require multiple inferences, this fact limits the potential of SNNs in terms of actual energy efficiency improvement. However, TEAS can achieve an average 22.9-fold improvement in energy efficiency compared to ANN models with similar parameters by significantly reducing the time steps.

[0068] In addition, this embodiment also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps in the above-mentioned time-adaptive spiking neural network accelerator optimization method. The above-mentioned time-adaptive spiking neural network accelerator optimization method has been described in detail and will not be elaborated here.

[0069] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to a computer program. The aforementioned computer program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps including those in the above method embodiments; and the aforementioned storage medium includes: various media such as ROM, RAM, magnetic disks, or optical discs that can store program codes.

[0070] This application embodiment also provides an electronic device. Figure 3 Shown is a schematic structural diagram of the electronic device 100 provided by this application embodiment. In some embodiments, the electronic device can be a mobile phone, a tablet computer, a wearable device, a vehicle-mounted device, an augmented reality (AR) / virtual reality (VR) device, a laptop computer, an ultra-mobile personal computer (UMPC), a netbook, a personal digital assistant (PDA), and other terminal devices. In addition, the multi-source heterogeneous digital twin integration method of the automatic loading system provided by this application can also be applied to databases, servers, and service response systems based on terminal artificial intelligence. This application embodiment does not impose any restrictions on the specific application scenarios of the multi-source heterogeneous digital twin integration method of the automatic loading system.

[0071] As Figure 3 shown, the electronic device 100 provided by this application embodiment includes a memory 101 and a processor 102.

[0072] The memory 101 is used to store a computer program; preferably, the memory 101 includes: various media such as ROM, RAM, magnetic disks, USB flash drives, memory cards, or optical discs that can store program codes.

[0073] Specifically, the memory 101 may include a computer system readable medium in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The electronic device 100 may further include other removable / non-removable, volatile / non-volatile computer system storage media. The memory 101 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of the present application.

[0074] The processor 102 is connected to the memory 101 and is configured to execute the computer program stored in the memory 101, so that the electronic device 100 executes the multi-source heterogeneous digital twin integration method of the automatic loading system provided in any embodiment of the present application.

[0075] Optionally, the processor 102 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0076] Optionally, the electronic device 100 in this embodiment may further include a display 103. The display 103 is communicatively connected to the memory 101 and the processor 102, and is configured to display a relevant GUI interaction interface for the multi-source heterogeneous digital twin integration method of the automatic loading system.

[0077] In summary, the present invention effectively evaluates the criticality of each layer in the SNN network by calculating the pulse activity of each layer of the computational network, helps to determine the contribution of each layer to the overall network performance, thereby guiding the optimization of the time step, and realizes the transmission of complete information between SNN layers with different time steps, effectively improving the computational efficiency and energy efficiency of the neural network accelerator. Therefore, the present invention effectively overcomes various disadvantages in the prior art and has high industrial utilization value.

[0078] The above embodiments are only illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Any person familiar with this technology can modify or change the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the technical field without departing from the spirit and technical idea disclosed by the present invention should still be covered by the claims of the present invention.

Claims

1. A method for optimizing a time - adaptive spiking neural network accelerator, characterized in that: The method includes: Calculating the pulse activity of a spiking neural network model in a spiking neural network accelerator to measure the contribution degree of each layer in the spiking neural network model to the spiking neural network model; Determining the loss function of the spiking neural network model based on the pulse activity; Performing iterative training on the spiking neural network model based on a data set and the loss function to determine corresponding time steps for each layer of the spiking neural network model; The spiking neural network accelerator executes an application program task based on the trained spiking neural network model.

2. The time - adaptive pulse neural network accelerator optimization method according to claim 1, wherein: The pulse activity is the ratio of the internal excitation of the current neuron to the length of its time step.

3. The time - adaptive pulse neural network accelerator optimization method according to claim 1, wherein: It further includes: Determining the transmitted pulse sequence for each layer based on the potential threshold of the excitation pulse and the excitation potential of the pulse.

4. The time - adaptive pulse neural network accelerator optimization method according to claim 3, wherein: For the spiking neural network model of the SNN-T type, the information passed to the next layer is a sequence of single spikes at time points where there is a single spike, where x' T is the spike generation time for compensation update in the SNN, x' T = x T + V th / V fire ; x T is the spike generation time in the SNN-T, V th is the potential threshold for spiking in the SNN, V fire is the average spiking potential of all spikes in the SNN-T; T l is the time step of the current layer, and T l+1 is the time step of the next layer.

5. The time - adaptive spiking neural network accelerator optimization method according to claim 3, wherein: For the SNN-R type of spiking neural network model, the information passed to the next layer is the sequence of spikes that occurred at the previous time points, where x′ R is the spike generation time for compensation update in the SNN, and x′ R = x R + 1 - V th / V′ fire , x R is the spike count in the SNN-R, V th is the potential threshold for firing spikes in the SNN, and V′ fire is the firing potential of the only spike in the SNN-R.

6. The time - adaptive pulse neural network accelerator optimization method according to claim 1, wherein: The loss function L of the spiking neural network model is: L = L CE + τl r ; Among them, Act(l) represents the pulse activity of the l-th layer, reflecting the impact of this layer on the network accuracy, Para(l) represents the total number of parameters of the l-th layer, reflecting the computational load of this layer, and L CE is the cross-entropy loss function for the original SNN training, τ is the penalty factor for balancing the cross-entropy loss and the regularization term loss, and l r is the regularization term loss function.

7. The time - adaptive pulse neural network accelerator optimization method according to claim 1, wherein: Determining the corresponding time steps for each layer of the spiking neural network model includes: In each training cycle, performing iterative processing on each layer and its respective time steps, and updating the membrane potential of the neuron based on the input pulse sequence and the cumulative input current; Updating the loss function and network weights; Updating the time steps of each layer.

8. The time-adaptive spiking neural network accelerator optimization method according to claim 1 or 7, characterized in that: The training process starts with the initialization of the time steps for each layer, which is dynamically set based on the pulse activity of each layer.

9. A computer storage medium storing program instructions, characterized in that: When the program instructions are executed, the steps of the time - adaptive spiking neural network accelerator optimization method according to any one of claims 1 to 8 are implemented.

10. An electronic device, characterized in that, The electronic device includes: A memory storing a computer program; A processor communicatively connected to the memory, and when calling the computer program, executes the time - adaptive spiking neural network accelerator optimization method according to any one of claims 1 to 8.

Images