Spiking Neural Network Training Accelerator for Early Image Termination

Solution Overview

Problem

Spiking neural networks face high energy consumption and prolonged training times due to repeated processes like inference, backpropagation, and weight update during training, especially with large datasets.

Innovation Solution

A supervised training accelerator for spiking neural networks that measures membrane potentials, extracts distribution data, calculates a threshold value, and early terminates training for images with no contribution using membrane potential differences.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If training is performed for all training images using the complete process (inference, backpropagation, and weight update), then training accuracy is improved, but training energy consumption and training time increase significantly

Engineering Contradiction:

Improvetraining accuracyVSAvoidtraining energy consumption

Core Design Contradiction:

ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent extracts and identifies images that do not contribute to training (non-contributory images) by analyzing membrane potential distributions. These non-contributory images are then excluded from the complete training process, removing unnecessary computational work while preserving training effectiveness. This is achieved by measuring membrane potentials during inference, comparing them against threshold values, and terminating training early for images that fail to meet the contribution criteria.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent performs preliminary analysis of membrane potential distributions before committing to full training for each image. By calculating threshold values based on membrane potential statistics and evaluating images against these thresholds in advance, the system determines which images warrant full training and which can be terminated early, thus preventing wasteful energy consumption before it occurs.

Inventive Principle:

Principle #10Preliminary action

2Reliability

If training is performed for all training images using the complete process (inference, backpropagation, and weight update), then training accuracy is improved, but training time increases significantly

Engineering Contradiction:

Improvetraining accuracyVSAvoidtraining time

Core Design Contradiction:

ReliabilityVSLoss of time

Solution Approach 1:

The patent extracts and identifies images that do not contribute to training (non-contributory images) by analyzing membrane potential distributions. These non-contributory images are then excluded from the complete training process, removing unnecessary computational work while preserving training effectiveness. This is achieved by measuring membrane potentials during inference, comparing them against threshold values, and terminating training early for images that fail to meet the contribution criteria.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent performs preliminary analysis of membrane potential distributions before committing to full training for each image. By calculating threshold values based on membrane potential statistics and evaluating images against these thresholds in advance, the system determines which images warrant full training and which can be terminated early, thus preventing wasteful energy consumption before it occurs.

Inventive Principle:

Principle #10Preliminary action

3Reliability

If the number of training images is increased to improve model generalization, then model performance is improved, but energy consumption and training time increase proportionally

Engineering Contradiction:

Improvemodel generalization performanceVSAvoidtraining efficiency

Core Design Contradiction:

ReliabilityVSProductivity

Solution Approach 1:

The patent extracts and identifies images that do not contribute to training (non-contributory images) by analyzing membrane potential distributions. These non-contributory images are then excluded from the complete training process, removing unnecessary computational work while preserving training effectiveness. This is achieved by measuring membrane potentials during inference, comparing them against threshold values, and terminating training early for images that fail to meet the contribution criteria.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the training parameter (number of images processed) dynamically by adjusting which images receive full training based on their contribution potential. By modifying the effective training set size through early termination of non-contributory images, the system maintains training efficiency even as the overall dataset size increases to improve model generalization.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS12632710B2Supervised training accelerator for temporal coding-based spiking neural network and operation method thereof

Publication Date: 2026.05.19 KOREA UNIV RES & BUSINESS FOUND

AI summary

Disclosed are a method for accelerating supervised training of a spiking neural network. The method includes measuring first and second membrane potentials for each time step during a training process, extracting distribution data of the first and second membrane potentials based on the first and second membrane potentials for the each time step, calculating a threshold value to be used in a subsequent training process based on the distribution data of the first and second membrane potentials, classifying images having no training contribution based on the threshold value calculated in a previous training process, and terminating the training at the time step based on determining that the image does not have the training contribution when a difference between the first and second membrane potentials in the time step is greater than the threshold value.