Asynchrony as a substitute for chaos in training ring spiking neural networks

Authors: Afroditi Talidou, Wilten Nicola

Presentation type: Flash talk at SNUFA 2025 online workshop (5-6 Nov 2025)

Abstract

Recurrent neural networks trained with the FORCE learning method typically rely on chaotic dynamics to provide the variability needed for learning complex temporal patterns. This perspective has been extended to spiking neural networks, where chaos is thought to be essential for generating the richness of internal dynamics that enables training. In this work, we challenge that assumption by studying spiking neural networks organized in ring topologies and propose that asynchrony, rather than chaos, is sufficient to drive FORCE learning. We support this claim by training ring networks of inhibitory interneurons on a range of dynamical systems tasks like stable oscillations, multi-stable states, and low-dimensional chaotic attractors. We show that asynchronous regimes provide enough diversity for effective learning by generating suitable temporal bases during the learning process. Our work demonstrates that asynchrony-based FORCE learning reproduces key biological features without reliance on high-dimensional chaotic states.