中央研究院 資訊科學研究所

This talk will introduce a new stochastic Markov recurrent neural network for deep sequential learning. This work aims to strengthen the learning capability in sequential prediction where the trajectory of sequential patterns is driven by a stochastic Markov process with the state transitions are carried out by a multiple-state long short-term memory model. Such a latent state machine is capable of identifying the complicated latent structure in heterogeneous sequential data. Gumbel-softmax is developed to implement the stochastic backpropagation algorithm with discrete states. Some extensions to structural learning, self-attention and multi-stream fusion will be discussed.

Jen-Tzung Chien is now with the Department of Electrical and Computer Engineering, National Chiao Tung University, Taiwan. His interests include machine learning, deep learning, natural language processing and computer vision.