Enhanced LSTM with Batch Normalization

Abstract

Recurrent neural networks (RNNs) are powerful models for sequence learning. However, the training of RNNs is complicated because the internal covariate shift problem, where the input distribution at each iteration changes during the training as the parameters have been updated. Although some work has applied batch normalization (BN) to alleviate this problem in long short-term memory (LSTM), unfortunately, BN has not been applied to the update of the LSTM cell. In this paper, to tackle the internal covariate shift problem of LSTM, we introduce a method to successfully integrate BN into the update of the LSTM cell. Experimental results on two benchmark data sets, i.e. MNIST and Fashion-MNIST, show that the proposed method, enhanced LSTM with BN (eLSTM-BN), has achieved a faster convergence than LSTM and its variants, while obtained higher classification accuracy on sequence learning tasks.