時系列データに対する掛け算レイヤー

抄録

In this paper, we propose a neural network layer which extracts discriminative higher-order auto-correlation features of the input data. We combine the proposed layer with models such as RNNs, LSTMs, and GRUs, and then apply it to sequence data classification. In the experiment, we investigate how the proposed layer works in the combined model. Furthermore, by comparing the difference between the proposed layer and a convolutional layer, we clarify the inherent characteristics of the proposed layer. Finally, by testing the combined model on different datasets, we find out appropriate applications.