Efficient Structured Inference for Stochastic Recurrent Neural Networks

Recent advances in sequential data modeling have suggested a class of models that combine recurrent neural networks with state space models. Despite the success, the huge model complexity has brought an important challenge to the corresponding inference methods. This paper introduces an structured inference algorithm to efficiently learn such models, including variants where the emission and transition distributions are modelled by deep neural networks. Our learning algorithm leverages a structured variational approximation parameterized by stochastic models and recurrent neural networks to approximate the posterior distribution. Experimental results on synthetic datasets have demonstrated the promising performance of the proposed method. In addition, our method has significantly outperformed the current state-of-the-art methods on music and speech modeling tasks.