Sparse-firing regularization methods for spiking neural networks with time-to-first-spike coding.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Spiking Neural Networks (SNNs) show promise for energy-efficient computation.
• Training SNNs with error backpropagation, particularly using firing times, is advancing.
• Time-to-first-spike (TTFS) coding enables low firing frequencies but its full potential at very low rates is unexplored.

Purpose of the Study:

• To investigate and enhance the information processing capabilities of TTFS-coded SNNs at reduced firing frequencies.
• To introduce novel regularization techniques for further decreasing firing rates in SNNs.
• To evaluate the effectiveness of these methods on standard image datasets.

Main Methods:

• Proposed two spike-timing-based sparse-firing (SSR) regularization methods.
• These methods utilize only firing timing and associated weights.
• Evaluated methods on MNIST, Fashion-MNIST, and CIFAR-10 datasets using MLP and CNN architectures.

Main Results:

• Demonstrated the effectiveness of SSR regularization in reducing firing frequencies of TTFS-coded SNNs.
• Investigated the impact of reduced firing rates on SNN performance across different datasets and network structures.
• Confirmed the feasibility of achieving lower firing frequencies without significant performance degradation.

Conclusions:

• The proposed SSR regularization methods effectively reduce the firing frequency in TTFS-coded SNNs.
• These techniques contribute to improving the energy efficiency of SNNs.
• Further research into low-frequency SNNs is warranted for advanced neuromorphic computing applications.