Learning compositional functions via multiplicative weight updates

TL;DR

This paper introduces Madam, a multiplicative optimizer inspired by biological synapses, which effectively trains neural networks without learning rate tuning and adapts to compressed models using logarithmic weight representations.

Contribution

It develops a new multiplicative optimization method, Madam, with theoretical guarantees for compositional functions, and demonstrates its effectiveness on neural networks and biological relevance.

Findings

Madam trains state-of-the-art neural networks without tuning

It adapts to compressed neural networks using logarithmic weights

Theoretical descent lemma supports multiplicative updates

Abstract

Compositionality is a basic structural feature of both biological and artificial neural networks. Learning compositional functions via gradient descent incurs well known problems like vanishing and exploding gradients, making careful learning rate tuning essential for real-world applications. This paper proves that multiplicative weight updates satisfy a descent lemma tailored to compositional functions. Based on this lemma, we derive Madam -- a multiplicative version of the Adam optimiser -- and show that it can train state of the art neural network architectures without learning rate tuning. We further show that Madam is easily adapted to train natively compressed neural networks by representing their weights in a logarithmic number system. We conclude by drawing connections between multiplicative weight updates and recent findings about synapses in biology.