Feature Overlapping: Temporal Differential Decoupling for Efficient Spiking Neural Network Training.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Spiking neural networks (SNNs) offer energy-efficient computation but face high training costs due to multi-time step processing.
• Existing methods for reducing SNN computational cost often focus on time steps without addressing temporal feature redundancy.

Purpose of the Study:

• To investigate and address the computational redundancy across temporal dimensions in SNNs.
• To propose a novel method for efficient SNN training and deployment.

Main Methods:

• Temporal differential decoupling (TDD) transforms network computation into the differential domain to separate static and dynamic features.
• The TDD-based differential domain low-sparsity approximation (TDD-DDLA) algorithm quantifies temporal feature contribution to gradient updates for energy optimization.
• Analysis of temporal feature evolution based on gradient sensitivity criterion.

Main Results:

• The proposed TDD framework significantly reduces redundant computations by disentangling temporal features.
• Achieved up to 80.9% fewer spikes per time step and 57.8% fewer total spikes.
• Maintained classification performance without degradation.

Conclusions:

• TDD offers a theoretically grounded approach to analyze and reduce temporal redundancy in SNNs.
• The method enables scalable, low-cost, and high-accuracy SNN deployment.
• This work provides a pathway for more efficient SNNs in practical applications.