Revisiting Activation Regularization for Language RNNs

TL;DR

This paper explores revisiting traditional regularization methods, like L2 and slowness regularization, for language RNNs, demonstrating they can improve performance with minimal architectural changes.

Contribution

It introduces the effective use of simple, traditional regularization techniques for RNNs, offering comparable or better results than complex methods.

Findings

L2 regularization improves language model performance.

Slowness regularization enhances RNN stability.

Techniques require minimal modifications to existing architectures.

Abstract

Recurrent neural networks (RNNs) serve as a fundamental building block for many sequence tasks across natural language processing. Recent research has focused on recurrent dropout techniques or custom RNN cells in order to improve performance. Both of these can require substantial modifications to the machine learning model or to the underlying RNN configurations. We revisit traditional regularization techniques, specifically L2 regularization on RNN activations and slowness regularization over successive hidden states, to improve the performance of RNNs on the task of language modeling. Both of these techniques require minimal modification to existing RNN architectures and result in performance improvements comparable or superior to more complicated regularization techniques or custom cell architectures. These regularization techniques can be used without any modification on optimized LSTM implementations such as the NVIDIA cuDNN LSTM.