Dynamic spatio-temporal pruning for efficient spiking neural networks

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking neural networks (SNNs) offer computational efficiency through event-driven processing and sparse data flow, mimicking biological neurons.
• Hardware constraints, particularly memory-processor bandwidth, necessitate efficient data handling in artificial neural networks (ANNs).
• Dynamical sparsity in SNNs minimizes data communication, reducing weight access and enabling effective pruning with less performance degradation compared to ANNs.

Purpose of the Study:

• To develop a spatio-temporal pruning algorithm for SNNs to reduce temporal redundancy in Dynamic Vision Sensor (DVS) datasets.
• To investigate the impact of pruning on SNN performance and parameter space reduction.
• To enhance the computational efficiency and memory demands of neuromorphic processors.

Main Methods:

• Proposed a spatio-temporal pruning algorithm that dynamically adapts to temporal redundancy in SNNs.
• Implemented spatial pruning based on global parameter statistics and inter-layer parameter counts.
• Conducted an ablation study to isolate and evaluate individual components of the proposed pruning approach.

Main Results:

• Achieved a 0.69% performance improvement on the DVS128 Gesture dataset, contrary to typical pruning expectations.
• Demonstrated an impressive 98.18% reduction in parameter space and a 50% reduction in time redundancy.
• The approach showed excellent performance across various datasets, particularly those with time-varying features.

Conclusions:

• The developed spatio-temporal pruning algorithm effectively reduces redundancy and parameter space in SNNs without compromising performance.
• This method offers a promising solution for efficient neuromorphic computing, especially for processing event-based sensory data.
• The findings highlight the potential of SNNs and advanced pruning techniques for next-generation AI hardware.