Dropframe Scheme in Recurrent Neural Networks for Time Series Modeling

Recurrent neural networks have been proven to be powerful models for many time series modeling applications, but are still limited by their tendency to overfit. The widely adopted dropout method, which is applied to feed-forward architectures, is shown to fail when applied to recurrent layers. We propose a novel dropframe scheme to allow arbitrary rearrangement of training data in RNNs. This scheme extends existing RNN models to skip frames randomly during the training phase, with the approximative behavior of data augmentation for time series. We apply this dropframe technique in long short-term memory (LSTM) and gated recurrent unit (GRU) models, assessing it on acoustic scene classification and video classification tasks. Our LSTM approach with dropframe scheme can perform significantly better than the vanilla LSTM architecture on LITIS Rouen dataset and DCASE2016 dataset, obtaining relative reductions of 4.7% and 3.6% on classification error.