BrainScale: Enabling Scalable Online Learning in Spiking Neural Networks

Abstract

Spiking neural networks (SNNs) represent a promising paradigm for understanding brain functions [1, 2] and developing neuromorphic intelligence [3, 4]. However, their potential remains largely unrealized due to a fundamental limitation: the lack of a scalable online learning system capable of supporting large-scale training of complex brain dynamics over long timescales. Current approaches are either limited to offline learning [5], impeding long-time task training; constrained by high memory complexity [6, 7], hindering network scaling; or using oversimplified models [8], failing to capture complex brain dynamics. Here, we introduce Brain-Scale, a system that synergizes model universality, computational efficiency, and engineering usability to enable scalable online learning in SNNs. First, BrainScale introduces standard model abstractions to support the training of diverse spiking networks. Second, BrainScale implements an online learning algorithm with linear memory complexity by exploiting intrinsic properties of SNN dynamics. Third, BrainScale provides an online learning compiler that automates the implementation of online learning for any user-defined models. Extensive evaluations across diverse SNN dynamics and computational tasks demonstrate that BrainScale consistently delivers strong training performance while maintaining extremely low memory usage and high computational efficiency. These properties enable the online training of a whole-brain-scale SNN of the Drosophila brain, accurately capturing functional activities across brain regions. These results position BrainScale as a foundational platform for advancing large-scale brain simulation and neuromorphic computing.