Log-Normal Multiplicative Dynamics for Stable Low-Precision Training of Large Networks
This addresses the challenge of energy-efficient hardware implementation for large networks by enabling low-precision training, representing a novel method rather than an incremental improvement.
The paper tackled the problem of training large neural networks with low-precision operations by proposing a Log-Normal Multiplicative Dynamics (LMD) algorithm inspired by biological synapses, achieving stable and accurate training-from-scratch for Vision Transformer and GPT-2 models.
Studies in neuroscience have shown that biological synapses follow a log-normal distribution whose transitioning can be explained by noisy multiplicative dynamics. Biological networks can function stably even under dynamically fluctuating conditions arising due to unreliable synaptic transmissions. Here we ask: Is it possible to design similar multiplicative training in artificial neural networks? To answer this question, we derive a Bayesian learning rule that assumes log-normal posterior distributions over weights which gives rise to a new Log-Normal Multiplicative Dynamics (LMD) algorithm. The algorithm uses multiplicative updates with both noise and regularization applied multiplicatively. The method is as easy to implement as Adam and only requires one additional vector to store. Our results show that LMD achieves stable and accurate training-from-scratch under low-precision forward operations for Vision Transformer and GPT-2. These results suggest that multiplicative dynamics, a biological feature, may enable stable low-precision inference and learning on future energy-efficient hardware.