Backpropagation With Sparsity Regularization for Spiking Neural Network Learning.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking Neural Networks (SNNs) offer a promising avenue for low-power, energy-efficient computing by leveraging biological system features.
• Existing SNN learning methods often struggle to optimize both spiking activity and synaptic connections effectively.

Purpose of the Study:

• To propose and evaluate a novel sparsity-driven learning algorithm for SNNs, termed backpropagation with sparsity regularization (BPSR).
• To enhance both spiking and synaptic sparsity in SNNs while maintaining or improving classification accuracy.
• To develop a learning algorithm that supports brain-like structure learning and suppresses information redundancy.

Main Methods:

• Developed BPSR, integrating backpropagation with spiking regularization to minimize firing rates and backpropagation with synaptic regularization for network rewiring.
• Implemented a rewiring mechanism based on weight and gradient to regulate synapse pruning and growth.
• Evaluated BPSR on diverse datasets including visual (MNIST, N-MNIST, CIFAR10) and sensor (MIT-BIH, gas sensor) data.

Main Results:

• The BPSR algorithm successfully achieved significant synaptic sparsity, creating networks structurally similar to biological systems.
• BPSR demonstrated a balance between classification accuracy and reduced spiking firing rates.
• The algorithm effectively suppressed information redundancy, facilitating more efficient SNN learning.
• Experimental results showed comparable or superior accuracy to existing methods on multiple benchmark datasets.

Conclusions:

• BPSR is an effective algorithm for training sparse and energy-efficient SNNs.
• The proposed method offers a viable approach for developing SNNs that closely mimic biological neural structures and functions.
• BPSR contributes to advancing SNNs for practical applications requiring high efficiency and accuracy.