Zeroth-Order Forward-Only SNN Training Inspiring Neuromorphic On-Chip Learning

The human brain is a biologically instantiated on-device neural system that integrates both learning and inference in a unified architecture, which enables rapid and flexible learning on-the-fly. This extraordinary online learning ability is realized through biological learning mechanisms operating on a well-initialized innate model. This work considers the on-chip edge learning upon pretrained models with zeroth-order (ZO) methods. ZO optimization methods, which resemble bio-plausible perturbation-based learning, offer a promising approach that enables learning with only forward passes and hence can significantly reduce the complexity of on-chip hardware implementation. However, in this work we show that applying ZO methods to spiking neural networks (SNNs) is non-trivial due to the step-function nature of spiking activation. We analyze the challenges posed by the spiking activation, and reveal a variance amplification effect of it. Based on this insight, we propose a subspace-based ZO (SZO) method that leverages the intrinsic low-dimensional structure of the SNN optimization trajectory. By learning in a low-dimensional subspace, SZO substantially enhances ZO learning efficacy, achieving accuracy comparable to first-order (FO) methods with faster learning speed than full-space BP. We evaluate SZO on model training from scratch, continual training, and unsupervised adaptation. Experimental results demonstrate that SZO closely approaches FO training performance for the first time while offering fast learning speed.

Future devices, such as smart terminals and robots, need to learn after deployment, without sending data to large cloud servers. Brain-inspired spiking neural networks are attractive for this goal because they communicate with brief electrical spikes, which can be efficient on chips. However, the standard way of training neural networks relies on a costly backward calculation that is difficult to implement on small neuromorphic hardware. A simpler alternative is to learn by only forward runs: slightly perturb the model, observe its output change, and update the model accordingly. We find that this method does not work well for spiking neural networks because their spike behavior makes the output change signal too noisy. To fix this, we show that spiking networks often learn along a much smaller set of important directions, and we restrict the perturbation in this compact subspace. Our method can train or adapt spiking networks using only forward computations, while reaching performance close to conventional training. This brings efficient on-chip learning closer to practical use in low-power, adaptive AI systems.