On the difficulty of training Recurrent Neural Networks

TLDR

There are two widely known issues with properly training Recurrent Neural Networks, the vanishing and the exploding gradient problems detailed in Bengio et al. (1994). The paper aims to deepen understanding of vanishing and exploding gradients in RNNs by analytically, geometrically, and dynamically exploring them, and proposes gradient norm clipping and a soft constraint to mitigate these issues. The authors analyze the problems using analytical, geometric, and dynamical systems methods, propose gradient norm clipping and a soft constraint, and empirically validate these solutions.

Abstract

There are two widely known issues with properly training Recurrent Neural Networks, the vanishing and the exploding gradient problems detailed in Bengio et al. (1994). In this paper we attempt to improve the understanding of the underlying issues by exploring these problems from an analytical, a geometric and a dynamical systems perspective. Our analysis is used to justify a simple yet effective solution. We propose a gradient norm clipping strategy to deal with exploding gradients and a soft constraint for the vanishing gradients problem. We validate empirically our hypothesis and proposed solutions in the experimental section.