Robust Spatiotemporal Prototype Learning for Spiking Neural Networks

Comprehensive Summary

This study proposes a spatiotemporal prototype (STP) learning for spiking neural networks (SNNs). Using learnable binarized prototypes for distance based decoding in this model combined with a co-training framework optimizes the model parameters and improves performance and robustness of SNNs. Existing methods such as network decoding, population coding, and time-to-first-spike (TTFS) decoding do not balance efficiency and performance. The STP-SNN model results showed comparable or surpassed accuracy on eight benchmark data sets (e.g., MNIST, CIFAR10-DVS, and SHD)

Outcomes and Implications

Effective decoding of electroencephalograms (EEG) using SNNs is important for diagnosis and research of different medical conditions. In brain-computer interfaces (BCI) and epilepsy detection, extracting temporal patterns is key and SNNs offer an advantage is temporal resolution. Research into robust and efficient SNNs could potentially make SNNs more suitable for wearable medical devices.