Integrating biological learning strategies into machine learning with plasticity-optimized neural networks.

Spike-timing--dependent plasticity (STDP) is a biological mechanism that adjusts the strength of synapses based on the timing, recurrence, and directionality of spikes between connected neurons. STDP is a critical learning mechanism in the mammalian brain, and researchers have recently investigated its potential to improve artificial neural network (ANN) performance. While biologically inspired STDP mechanisms have been widely used in 3rd-generation spiking neural networks (SNNs), STDP is currently underutilized in 2nd-generation classical neural networks due to a lack of accessible tools for incorporating it into ANN training mechanisms. In this study, we present a Pytorch implementation of STDP as an optimizer and extensively characterize the behavior and performance of both 2nd- and 3rd-generation NNs with respect to STDP parameters. Our study shows that STDP can be integrated into machine learning workflows to enhance performance by strengthening connections based on coordinated activations during training. Our Pytorch implementation of STDP can help bridge the gap between machine learning and biological simulations, enabling the exploration of biological learning strategies.