Jujubecake: An Extension of LSTM Considering Correlation among Input Blocks

This paper proposes a new neural network unit for sequential input, called Jujubecake Unit or Jujubecake Cell. In the chain structure formed by this unit, not only the correlation among individual inputs but also the correlation among blocks consist of several inputs is considered so that that performance can be improved. In this paper, Jujubecake model and the multi-layer LSTM model are evaluated in two experiments: language modelling and audio event detection. The experimental results show that Jujubecake Unit proposed in this paper outperforms the traditional LSTM.