Early judgment learning based on timestep splitting of spiking neural network and operation method thereof

Abstract

The present invention relates to an early decision learning device based on time step segmentation of a spiking neural network and a method of operation thereof. In one embodiment of the present invention, the early decision learning acceleration method based on time step segmentation may include a time step segmentation step for segmenting a time step; a membrane potential measurement step for measuring first and second membrane potentials for each segmented time step during a current learning process; a threshold value calculation step for calculating a threshold value to be used in the next learning process based on the first and second membrane potentials; and an early learning termination step for terminating learning in the segmented time step by determining that there is no learning contribution to the image when the difference between the first and second membrane potentials in the segmented time step is greater than the threshold value.

Description

Technology Field

[0001] The present invention relates to an early decision learning device based on time-step segmentation of a spiking neural network and a method of operation thereof. Background Technology

[0002] Spiking neural networks receive and transmit information in spike units of 0s and 1s, offering the advantage of low power consumption in hardware implementation. Additionally, the introduction of temporal coding reduces the number of spikes in the inference process, significantly decreasing the energy required for the inference process. However, there is a problem in that the training of neural networks consumes a significant amount of energy and time because the training process of inference, backpropagation, and weight updates is repeated over a very large number of training images. The problem to be solved

[0003] The present invention aims to provide a time step segmentation-based early judgment learning method that can reduce learning energy and time by segmenting time steps and determining the network contribution of training images, thereby terminating the learning of images with no learning contribution early. means of solving the problem

[0004] In one embodiment of the present invention, the time step segmentation-based early judgment learning acceleration method may include a time step segmentation step for segmenting a time step; a membrane potential measurement step for measuring a first and second membrane potential for each segmented time step during a current learning process; a threshold value calculation step for calculating a threshold value to be used in the next learning process based on the first and second membrane potentials; and an early learning termination step for determining that there is no learning contribution to the image and terminating learning in the segmented time step when the difference between the first and second membrane potentials in the segmented time step is greater than the threshold value.

[0005] In one embodiment of the present invention, the first membrane potential is the membrane potential of the correct answer neuron, and the second membrane potential may be the largest membrane potential among a plurality of membrane potentials, excluding the membrane potential value of the correct answer neuron.

[0006] In one embodiment of the present invention, the threshold value calculation step calculates the threshold value based on Equation 1, wherein Equation 1 is a=m+2σ, where a is the threshold value, m is the average of the membrane potential difference distribution of the image with the learning contribution, and σ is the deviation of the membrane potential difference distribution of the image with the learning contribution.

[0007] In one embodiment of the present invention, the learning early termination step sets the boundary between the image without learning contribution and the image with learning contribution based on Equation 2, wherein Equation 2 is y=xa, where a is a threshold value calculated in the previous learning process, x is the membrane potential of the correct neuron, and y is the largest membrane potential among the membrane potentials excluding the membrane potential value of the correct neuron among the plurality of membrane potentials.

[0008] In one embodiment of the present invention, the learning early termination step determines an image without a learning contribution based on Equation 3 modified from Equation 2, wherein Equation 3 is xy≥a, where a is a threshold value calculated in the previous learning process, where x is the membrane potential of the correct neuron, and where y is the largest membrane potential among the membrane potentials excluding the membrane potential value of the correct neuron among the plurality of membrane potentials.

[0009] In one embodiment of the present invention, the learning early termination step may terminate learning at the subdivided time step by substituting the first and second membrane potentials and the threshold value for each subdivided time step into the mathematical formula 3 and determining that the image has no learning contribution when the mathematical formula 3 is satisfied, and proceed with learning by determining that the image has a learning contribution when the first and second membrane potentials and the threshold value for each subdivided time step are substituting the first and second membrane potentials and the threshold value for each subdivided time step into the mathematical formula 3 and determining that the image has a learning contribution when the mathematical formula 3 is not satisfied.

[0010] In one embodiment of the present invention, the subdivided time step may be a time step in which one time step is subdivided into 2 to 16.

[0011] An early judgment learning acceleration device based on time step segmentation according to one embodiment of the present invention may include an input layer module into which an input spike signal of a spiking neural network is input, a hidden layer module that receives the input spike signal, an output layer module that receives the input spike signal from the hidden layer module, determines an image without a learning contribution, and calculates a threshold value for determining the presence or absence of the learning contribution, and a global controller that terminates the learning process based on the presence or absence of the learning contribution determined by the output layer module.

[0012] In one embodiment of the present invention, the hidden layer module may include a membrane potential update module that receives an input spike signal from the input layer module and calculates the membrane potential of a plurality of neurons based on the input spike, a weight update module that adds weights of the input spikes to the membrane potential based on the input of the input spike signal, a membrane potential buffer that stores the membrane potential of the plurality of neurons, and a spike time buffer that stores the spike occurrence time of the plurality of neurons.

[0013] In one embodiment of the present invention, the output layer module may include: a spike buffer that receives the input spike signal from the hidden layer module; a membrane potential update module that receives the input spike signal from the spike buffer and calculates the membrane potential of a plurality of neurons based on the input spike; a weight update module that adds weights of the input spikes to the membrane potential based on the input of the input spike signal; a membrane potential buffer that stores the membrane potential of the plurality of neurons; a spike time buffer that stores the spike occurrence times of the plurality of neurons; an error calculation unit that calculates the difference between the spike occurrence times of the plurality of neurons and the target correct answer signal after the forward propagation process is completely finished during the learning process; an early determination unit that determines an image without the learning contribution when the operation of each time step in the forward propagation process is finished; and a threshold calculation unit that calculates a threshold value to be used in the next learning after the forward propagation process is completely finished.

[0014] In one embodiment of the present invention, the early determination unit may determine that there is no learning contribution to the image when the difference between the first and second membrane potentials at the time step is greater than the threshold value, and terminate the learning at the time step.

[0015] In one embodiment of the present invention, the threshold value calculation unit can calculate the threshold value to be used in the next learning process based on distribution data of the first and second membrane potentials among the membrane potentials of the plurality of neurons.

[0016] In one embodiment of the present invention, the first membrane potential is the membrane potential of the correct answer neuron, and the second membrane potential may be the largest membrane potential among a plurality of membrane potentials, excluding the membrane potential value of the correct answer neuron. Effects of the invention

[0017] The time step segmentation-based early judgment learning method of the present invention can reduce learning energy and time by segmenting time steps and determining the network contribution of training images, thereby terminating the learning of images with no learning contribution early. Brief explanation of the drawing

[0018] Figure 1 is a graph showing the proportion of images with no learning contribution according to the learning process. FIG. 2 is a flowchart showing an early decision learning acceleration method based on time step segmentation according to an embodiment of the present invention. FIG. 3 is a figure showing an example of a time step subdivision step according to one embodiment of the present invention. FIG. 4 is a diagram showing a method for calculating a threshold value according to an embodiment of the present invention. FIG. 5 is a figure showing a method for early termination of learning by determining an image with no learning contribution according to an embodiment of the present invention. FIG. 6 is a figure showing an example of early termination of learning for a subdivided time step according to one embodiment of the present invention. FIG. 7 is a figure showing an early judgment learning acceleration device based on time step segmentation according to an embodiment of the present invention. Specific details for implementing the invention

[0019] Hereinafter, in order to explain the technical concept of the present application in detail so that a person skilled in the art can easily implement it, embodiments of the present application will be described with reference to the attached drawings.

[0020] Figure 1 is a graph showing the proportion of images with no learning contribution according to the learning process.

[0021] Referring to Fig. 1, the proportion of images that do not contribute to training during the training process can be measured using multiple datasets. In this case, the x-axis represents the training process, and the y-axis represents the proportion of images that do not contribute to training among the training data for each of the multiple datasets. The multiple datasets may be MNIST, Fashion-MNIST, and ETH-80.

[0022] In the case of the MNIST dataset, if the size of the training set is 60,000 and the proportion of images with no training contribution during the training process is 0.5, there may be 30,000 images with no training contribution.

[0023] For the Fashion-MNIST dataset, if the size of the training set is 60,000 and the proportion of images with no training contribution during the training process is 0.5, there may be 30,000 images with no training contribution.

[0024] The ETH-80 dataset consists of a total of 3,280 images. For example, 2,624 images, or 80% of the total, can be used as the training set, and 656 images, or 20%, can be used as the inference set.

[0025] As mentioned above, the training process of spiking neural networks based on conventional technology is characterized by a high proportion of non-contributing images, and it can be observed that the amount of non-contributing images increases further as training progresses. Consequently, when the amount of non-contributing images increases during the training process, there is a problem in that energy consumption increases and computation time becomes prolonged.

[0026] FIG. 2 is a flowchart showing an early decision learning acceleration method based on time step segmentation according to an embodiment of the present invention.

[0027] In step S10, the time step segmentation-based early decision learning acceleration device can segment the time steps. For example, the time step segmentation-based early decision learning acceleration device can segment one time step into 2 to 16.

[0028] In step S20, the time step segmentation-based early judgment learning acceleration device can measure membrane potential. For example, the time step segmentation-based early judgment learning acceleration device can measure the first and second membrane potentials for each segmented time step during the current learning process.

[0029] In this case, the first membrane potential is the membrane potential of the correct neuron, and the second membrane potential may be the largest membrane potential among the multiple membrane potentials, excluding the membrane potential value of the correct neuron.

[0030] In step S30, the time step segmentation-based early judgment learning acceleration device can calculate a threshold value. For example, the time step segmentation-based early judgment learning acceleration device can calculate a threshold value to be used in the next learning process based on the distribution data of the current first and second membrane potentials. At this time, the threshold value can be calculated based on Equation 1, and specific details regarding this will be described later in FIG. 4.

[0031] In step S40, the time step segmentation-based early judgment learning acceleration device may determine that there is no learning contribution to the image and terminate learning in the segmented time step if the difference between the first and second membrane potentials in the segmented time step is greater than the threshold value. Accordingly, by terminating learning for images with no learning contribution, the time step segmentation-based early judgment learning acceleration device can reduce the amount of computation and computation time without reducing accuracy.

[0032] More specifically, the time step segmentation-based early decision learning acceleration device can reduce computational load and computation time compared to terminating training for images with no training contribution at a single time step by segmenting a single time step and terminating training for images with no training contribution.

[0033] FIG. 3 is a figure showing an example of a time step subdivision step according to one embodiment of the present invention.

[0034] Referring to FIG. 3, one time step consists of 16 cells, and at least one of the 16 cells may have an input spike. For example, as shown in FIG. 3, one time step includes the 1st to 16th cells, and the 1st, 4th, 6th, 11th, and 13th cells may have input spikes.

[0035] The time step segmentation-based early decision learning acceleration device can subdivide one time step into four.

[0036] At this time, the first subdivision area may include the 1st, 3rd, 9th, and 11th cells, the second subdivision area may include the 2nd, 4th, 10th, and 12th cells, the third subdivision area may include the 5th, 7th, 13th, and 15th cells, and the fourth subdivision area may include the 6th, 8th, 14th, and 16th cells.

[0037] However, the above description is merely an example, and the time steps can be subdivided into 2 to 16 steps.

[0038] For example, when a time step is subdivided into two time steps, the first subdivision area may include the 1st to 8th cells, and the second subdivision area may include the 9th to 16th cells.

[0039] As another example, when a time step is subdivided into two time steps, the first subdivision area may include the 1st, 3rd, 5th, 7th, 9th, 11th, 13th, and 15th cells, and the second subdivision area may include the 2nd, 4th, 6th, 8th, 10th, 12th, 14th, and 16th cells.

[0040] In an embodiment of the present application, the designer may appropriately design the number of time step subdivisions according to the hardware configuration. For example, increasing the number of time step subdivisions may enable more early judgments to be performed, thereby increasing the learning acceleration effect. However, as the number of time step subdivisions increases, the hardware may incur additional energy consumption because the early judgment module must operate more frequently. Therefore, the designer may appropriately design the number of time steps by considering the hardware configuration and performance.

[0041] FIG. 4 is a diagram showing a method for calculating a threshold value according to an embodiment of the present invention.

[0042] Referring to Fig. 4, the time-step segmentation-based early decision learning acceleration device can represent images with a learning contribution (NZL images) and images without a learning contribution (ZL images) in the form of a Gaussian during the learning process.

[0043] For example, a time-step segmentation-based early decision learning acceleration device can represent images with a learning contribution (NZL images) and images without a learning contribution (ZL images) in the form of a Gaussian where the x-axis is the membrane potential gap and the y-axis is the number of images with the corresponding membrane potential gap.

[0044] A time-step segmentation-based early decision learning acceleration device can calculate a threshold based on data represented in a Gaussian form, based on the gap in membrane potential according to the time step and the number of images having the corresponding membrane potential gap. In this case, the gap in membrane potential may be the difference between the membrane potential of the correct neuron and the maximum value among the membrane potentials of the remaining neurons.

[0045] The threshold value can be calculated based on Equation 1. Equation 1 is as follows.

[0046] (Mathematical Formula 1)

[0047] Here, a is the threshold, m is the mean of the membrane potential difference distribution of the images with the learning contribution, and σ may be the deviation of the membrane potential difference distribution of the images with the learning contribution.

[0048] Since the threshold is calculated for each time step and learning process rather than being found empirically, it can be calculated automatically regardless of the datanet and network.

[0049] FIG. 5 is a figure showing a method for early termination of learning by determining an image with no learning contribution according to an embodiment of the present invention.

[0050] Referring to Fig. 5, the time step segmentation-based early decision learning acceleration device can represent the distribution of images with no learning contribution and images with a learning contribution in the form of a scatter plot where the x-axis is the membrane potential of the correct neuron (Vans) and the y-axis is the membrane potential of the maximum value among the membrane potentials of neurons excluding the correct neuron (Vres,max).

[0051] The time-step segmentation-based early judgment learning acceleration device can determine the relationship between the first and second membrane potentials in the training images during one training iteration of the learning process according to the time step, based on data represented in the form of a scatter plot.

[0052] The time-step segmentation-based early judgment learning acceleration device can set the boundary between images with no learning contribution and images with a learning contribution based on Equation 2.

[0053] Mathematical equation 2 can be y=xa.

[0054] At this time, a is a threshold value calculated in the previous learning process, x is the first membrane potential which is the membrane potential of the correct neuron, and y may be the second membrane potential which is the largest membrane potential among the multiple membrane potentials excluding the membrane potential value of the correct neuron.

[0055] The time-step segmentation-based early decision learning acceleration device can terminate learning for images with no learning contribution based on Equation 3, which is generated by transforming Equation 2.

[0056] Mathematical equation 3 can be xy≥a.

[0057] At this time, a is a threshold value calculated in the previous learning process, x is the first membrane potential which is the membrane potential of the correct neuron, and y may be the second membrane potential which is the largest membrane potential among the multiple membrane potentials excluding the membrane potential value of the correct neuron.

[0058] The time step segmentation-based early decision learning acceleration device can terminate ongoing learning by determining that the image has no learning contribution when Equation 3 is satisfied. On the other hand, the time step segmentation-based early decision learning acceleration device can proceed with learning by determining that the image has a learning contribution when Equation 3 is not satisfied.

[0059] That is, when the difference between the first membrane potential and the second membrane potential is greater than the threshold value (the right side of the line), the time-step segmentation-based early judgment learning acceleration device determines that there is no learning contribution from the image and can terminate the learning.

[0060] FIG. 6 is a figure showing an example of early termination of learning for a subdivided time step according to one embodiment of the present invention.

[0061] Referring to FIG. 6, the time step segmentation-based early decision learning acceleration device can terminate learning at a segmented time step when the difference (Vgap) between the first and second membrane potentials is greater than or equal to a threshold value at that segmented time step.

[0062] For example, a time step is subdivided into first to fourth time steps (T1, T2, T3, T4), and if the difference (Vgap) between the first and second membrane potentials between the first and second time steps (T1, T2) is greater than or equal to a threshold value, the time step subdivision-based early judgment learning acceleration device can terminate learning at the second time step (T2).

[0063] As described above, the time step segmentation-based early decision learning acceleration device may terminate learning in the middle of a time step based on the first and second membrane potential difference (Vgap) and the threshold value (a), rather than terminating learning for the entire time step. Accordingly, the amount of computation and computation time for learning can be reduced.

[0064] FIG. 7 is a figure showing an early judgment learning acceleration device based on time step segmentation according to one embodiment of the present invention.

[0065] Referring to FIG. 7, the time step segmentation-based early decision learning acceleration device (10) may include an input layer module (200), a hidden layer module (300), an output layer module (400), and a global controller (500).

[0066] The input layer module (200) can receive input spike signals of the spiking neural network.

[0067] The hidden layer module (300) can receive an input spike signal from the input layer module (200). At this time, the hidden layer module (300) may include a membrane potential update module (310), a weight update module (320), a membrane potential buffer (330), and a spike time buffer (340).

[0068] The membrane potential update module (310) receives an input spike signal from the input layer module (200) and can calculate the membrane potential of multiple neurons based on the input spike.

[0069] The membrane potential update module (310) may include a learning weight storage unit (311) and an inference weight storage unit (312).

[0070] The learning weight storage unit (311) can receive weights from the weight update module (320) during the learning process and update the membrane potential.

[0071] The inference weight storage unit (312) can update the membrane potential by receiving weights from the weight update module (320) during the inference process.

[0072] The weight update module (320) can add weights of input spikes to the membrane potential based on the input of the input spike signal. The weight update module (320) can transmit weight information to the learning weight storage unit (311) and the inference weight storage unit (312). Additionally, the weight update module (320) can transmit weight information to the weight update module (430) of the output layer module (400).

[0073] The membrane potential buffer (330) can store the membrane potentials of multiple neurons. The membrane potential buffer can transmit or receive membrane potential values ​​to the membrane potential update module (310). Additionally, the membrane potential buffer (330) can transmit information about the membrane potential to the spike time buffer (340).

[0074] The spike time buffer (340) can store the spike occurrence times of multiple neurons.

[0075] The output layer module (400) receives an input spike signal from the hidden layer module (300), determines an image without a learning contribution, and can calculate a threshold value to determine the presence or absence of a learning contribution. At this time, the output layer module (400) may include a spike buffer (410), a membrane potential update module (420), a weight update module (430), a membrane potential buffer (440), an early determination unit (450), a threshold value calculation unit (460), a spike time buffer (470), and an error calculation unit (480).

[0076] The spike buffer (410) can receive an input spike signal from the membrane potential update module of the hidden layer module (300). The spike buffer (410) can transmit the input spike signal to the membrane potential update module (420).

[0077] The membrane potential update module (420) receives an input spike signal from the hidden layer module (300) and can calculate the membrane potential of multiple neurons based on the input spike.

[0078] The membrane potential update module (420) may include a learning weight storage unit (421) and an inference weight storage unit (422).

[0079] The learning weight storage unit (421) can receive weights from the weight update module (430) during the learning process and update the membrane potential.

[0080] The inference weight storage unit (422) can update the membrane potential by receiving weights from the weight update module (430) during the inference process.

[0081] The weight update module (430) can add weights of input spikes to the membrane potential based on the input of the input spike signal. The weight update module (430) can transmit weight information to the learning weight storage unit (421) and the inference weight storage unit (422). Additionally, the weight update module (430) can transmit weight information to the weight update module (320) of the hidden layer module (300).

[0082] The membrane potential buffer (440) can store the membrane potentials of multiple neurons. The membrane potential buffer (440) can transmit or receive membrane potential values ​​to the membrane potential update module (420). Additionally, the membrane potential buffer (440) can transmit information about the membrane potential to the spike time buffer (470) and the early determination unit (450).

[0083] The early judgment unit (450) can determine an image that has no learning contribution when the operation of each time step is finished during the forward propagation process. The early judgment unit (450) can transmit the judgment result to the global controller (500).

[0084] For example, if the difference between the first and second membrane potentials at a time step is greater than a threshold value, the early judgment unit (450) may determine that there is no learning contribution to the image and terminate learning at the time step. At this time, the first membrane potential is the membrane potential of the correct answer neuron, and the second membrane potential may be the largest membrane potential among the membrane potentials excluding the membrane potential of the correct answer neuron among the plurality of membrane potentials.

[0085] The threshold calculation unit (460) can calculate the threshold to be used in the next learning after the entire forward propagation process is completed. Additionally, the threshold calculation unit (460) can transmit the threshold calculated in the previous learning to the early decision unit (450).

[0086] The spike time buffer (470) can store the spike occurrence times of multiple neurons. The spike time buffer (470) can transmit the spike occurrence times to the error calculation unit (480).

[0087] The error calculation unit (480) can calculate whether an error exists by calculating the difference between the spike occurrence times of multiple neurons after the entire forward propagation process is completed and the target correct answer signal. The error calculation unit can transmit the calculation result to the weight update module (430).

[0088] The global controller (500) can terminate the learning process based on whether there is a learning contribution determined by the early determination unit (450) of the output layer module (400). For example, if the early determination unit (450) determines that the image has no learning contribution, the global controller (500) can terminate the learning process.

[0089] The embodiments disclosed in this specification and drawings are provided merely as specific examples to facilitate the explanation and understanding of the contents of this disclosure and are not intended to limit the scope of this disclosure. Accordingly, the scope of this disclosure should be interpreted to include all modifications or variations derived based on the technical concept of this disclosure, in addition to the embodiments disclosed herein. Explanation of the symbols

[0090] 10: Time-step segmentation-based early decision learning accelerator 200 : Input layer module 300 : Hidden Layer Module 310, 420 : Membrane potential update module 311, 421: Learning weight storage 312, 422: Inference weight storage 320, 430: Weight update module 330, 440: Membrane potential buffer 340, 470 : Spike time buffer 400 : Output layer module 410 : Spike buffer 450 : Early Judgment Unit 460 : Threshold calculation unit 480 : Error Calculation Unit 500 : Global Controller

Claims

Claim 1 A method for accelerating early judgment learning based on time step segmentation, performed in a learning acceleration device of a spiking neural network, comprising: a time step segmentation step in which the learning acceleration device segments a time step; a membrane potential measurement step in which the learning acceleration device measures, for each segmented time step during the current learning process, a first membrane potential corresponding to the membrane potential of a correct answer neuron and a second membrane potential which is the largest membrane potential among a plurality of membrane potentials excluding the membrane potential value of the correct answer neuron; a threshold value calculation step in which the learning acceleration device automatically calculates a threshold value to be used in the next learning process based on the first and second membrane potentials; and an early learning termination step in which the learning acceleration device determines that there is no learning contribution of the spiking neural network to the image used for learning the spiking neural network and terminates learning within the segmented time step when the difference between the first and second membrane potentials within the segmented time step is greater than the threshold value. Claim 2 delete Claim 3 A time-step segmentation-based early judgment learning acceleration method, wherein the threshold calculation step calculates the threshold based on Equation 1, wherein Equation 1 is a=m+2σ, where a is the threshold, m is the average of the membrane potential difference distribution of the image with the learning contribution, and σ is the deviation of the membrane potential difference distribution of the image with the learning contribution. Claim 4 In claim 1, the learning early termination step sets the boundary between the image without learning contribution and the image with learning contribution based on Equation 2, wherein Equation 2 is y=xa, where a is a threshold value calculated in the previous learning process, x is the membrane potential of the correct answer neuron, and y is the largest membrane potential among the membrane potentials excluding the membrane potential value of the correct answer neuron among the plurality of membrane potentials. Claim 5 In claim 4, the learning early termination step determines an image without a learning contribution based on Equation 3 modified from Equation 2, wherein Equation 3 is xy≥a, where a is a threshold value calculated in the previous learning process, where x is the membrane potential of the correct answer neuron, and where y is the largest membrane potential among the membrane potentials excluding the membrane potential value of the correct answer neuron among the plurality of membrane potentials. Claim 6 In claim 5, the learning early termination step is a method for accelerating learning based on time step segmentation, wherein the first and second membrane potentials and the threshold value for each segmented time step are substituted into Equation 3, and if Equation 3 is satisfied, the image is determined to have no learning contribution and the learning is terminated at the segmented time step, and if Equation 3 is not satisfied, the first and second membrane potentials and the threshold value for each segmented time step are substituted into Equation 3, the image is determined to have a learning contribution and the learning is proceeded. Claim 7 A method for accelerating early decision learning based on time step subdivision according to claim 1, wherein the subdivided time step is a time step in which one time step is subdivided into 2 to 16. Claim 8 An early decision learning acceleration device based on time step segmentation, comprising: an input layer module into which an input spike signal of a spiking neural network is input; a hidden layer module including a membrane potential update module that receives the input spike signal and calculates the membrane potentials of a plurality of neurons for each segmented time step during the current learning process based on the input spike signal; an output layer module that receives the input spike signal from the hidden layer module, determines an image used for learning the spiking neural network that has no learning contribution of the spiking neural network within the segmented time step, and automatically calculates a threshold value for determining the presence or absence of the learning contribution; and a global controller that terminates the learning process within the segmented time step based on the presence or absence of the learning contribution determined by the output layer module. Claim 9 In claim 8, the hidden layer module comprises: a weight update module that adds weights of input spikes based on the input of the input spike signal to the membrane potential; a membrane potential buffer that stores the membrane potentials of the plurality of neurons; and a spike time buffer that stores the spike occurrence times of the plurality of neurons, a time step segmentation-based early decision learning acceleration device. Claim 10 In claim 8, the output layer module comprises: a spike buffer that receives the input spike signal from the hidden layer module; a membrane potential update module that receives the input spike signal from the spike buffer and calculates the membrane potential of a plurality of neurons based on the input spikes; a weight update module that adds weights of the input spikes to the membrane potential based on the input of the input spike signal; a membrane potential buffer that stores the membrane potential of the plurality of neurons; a spike time buffer that stores the spike occurrence times of the plurality of neurons; an error calculation unit that calculates the difference between the spike occurrence times of the plurality of neurons and the target correct answer signal after the forward propagation process is completely finished during the learning process; an early judgment unit that determines an image without the learning contribution when the operation of each time step in the forward propagation process is finished; and a threshold calculation unit that calculates a threshold value to be used in the next learning after the forward propagation process is completely finished, a time step segmentation-based early judgment learning acceleration device. Claim 11 In claim 10, the above-mentioned early judgment unit determines that there is no learning contribution to the image and terminates learning at the above-mentioned time step when the difference between the first and second membrane potentials at the time step is greater than the above-mentioned threshold value, thereby forming a time step segmentation-based early judgment learning acceleration device. Claim 12 In claim 10, the threshold calculation unit calculates the threshold value to be used in the next learning process based on distribution data of the first and second membrane potentials among the membrane potentials of the plurality of neurons, a time-step segmentation-based early judgment learning acceleration device. Claim 13 In claim 12, the first membrane potential is the membrane potential of the correct answer neuron, and the second membrane potential is the largest membrane potential among the membrane potentials excluding the membrane potential value of the correct answer neuron among a plurality of membrane potentials.

Images