On the biological plausibility of orthogonal initialisation for solving gradient instability in deep neural networks

TL;DR

This paper introduces biologically plausible methods for orthogonal weight initialization in neural networks, demonstrating their theoretical convergence and empirical superiority over random initializations in addressing gradient instability.

Contribution

It proposes two biologically plausible schemes for orthogonal weight initialization, with theoretical guarantees and improved empirical performance.

Findings

Pre-training orthogonalisation always converges.

Proposed schemes outperform random initializations.

Methods are biologically plausible.

Abstract

Initialising the synaptic weights of artificial neural networks (ANNs) with orthogonal matrices is known to alleviate vanishing and exploding gradient problems. A major objection against such initialisation schemes is that they are deemed biologically implausible as they mandate factorization techniques that are difficult to attribute to a neurobiological process. This paper presents two initialisation schemes that allow a network to naturally evolve its weights to form orthogonal matrices, provides theoretical analysis that pre-training orthogonalisation always converges, and empirically confirms that the proposed schemes outperform randomly initialised recurrent and feedforward networks.