On the difficulty of training Recurrent Neural Networks

TL;DR

This paper analyzes the challenges of training RNNs, specifically vanishing and exploding gradients, from multiple perspectives and proposes practical solutions like gradient clipping and soft constraints, validated through experiments.

Contribution

It offers a deeper understanding of RNN training issues and introduces simple, effective methods to mitigate vanishing and exploding gradients.

Findings

Gradient norm clipping effectively handles exploding gradients.

Soft constraints help mitigate vanishing gradients.

Empirical validation confirms the proposed solutions improve RNN training.

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.