Robust Spatiotemporal Prototype Learning for Spiking Neural Networks.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Spiking neural networks (SNNs) offer energy efficiency advantages over traditional artificial neural networks (ANNs).
• The spiking decoder is critical for SNN performance, but current decoding methods lack robustness and suitable training frameworks.
• Existing alternatives to rate coding often result in diminished overall performance.

Purpose of the Study:

• To introduce a novel decoding method for SNNs that enhances robustness and performance.
• To develop a cotraining framework for joint optimization of SNN prototypes and model parameters.

Main Methods:

• Proposed spatiotemporal prototype (STP) learning using multiple learnable binarized prototypes for distance-based decoding.
• Introduced a cotraining framework for mutual adaptation of prototypes and model parameters.
• Employed supervised learning to cluster feature centers around prototypes while ensuring inter-prototype spacing for noise resilience.

Main Results:

• The STP-SNN model achieved performance comparable to or exceeding state-of-the-art methods on eight diverse benchmark datasets.
• Demonstrated exceptional robustness and stability in multitask experiments.
• STP learning effectively clusters feature centers and maintains prototype separation, enhancing stability.

Conclusions:

• Spatiotemporal prototype (STP) learning is an effective strategy for improving the performance and robustness of spiking neural networks.
• The proposed cotraining framework facilitates mutual adaptation, leading to superior stability.
• STP learning addresses key limitations in SNN decoding, paving the way for more reliable SNN applications.