First-spike coding promotes accurate and efficient spiking neural networks for discrete events with rich temporal...

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking neural networks (SNNs) excel at processing event-based data.
• Traditional SNNs often use rate-coding (FR), neglecting precise spike timing.
• Temporal coding, like time-to-first-spike (TTFS), is efficient but challenging to train, especially for real-world event data due to unrealistic constraints.

Purpose of the Study:

• To introduce a novel first-spike (FS) coding strategy for SNNs to classify real-world event sequences.
• To investigate the significance of first-spike timing in complex temporal data.
• To develop a robust training method for FS coding in SNNs.

Main Methods:

• Proposed a novel surrogate gradient learning method for discrete spike trains to enable FS coding.
• Implemented a forward pass encoding discrete spike times into FS times.
• Developed an error assignment method using a Gaussian window for backpropagation and supervised learning for spike trains.

Main Results:

• FS coding achieved accuracy comparable to FR coding.
• FS coding demonstrated superior energy efficiency.
• FS coding revealed distinct neuronal dynamics, particularly on data with rich temporal structures.
• A longer time delay in the first spike correlated with increased classification accuracy.

Conclusions:

• First-spike (FS) coding is a viable and efficient alternative to rate-coding (FR) for SNNs processing complex temporal event data.
• The timing of the first spike carries significant information crucial for accurate classification.
• The developed surrogate gradient learning method effectively trains FS coding in SNNs.