Deep Generative Networks For Sequence Prediction

TL;DR

This thesis explores deep generative models based on GSNs for unsupervised sequence prediction, demonstrating their effectiveness on diverse high-dimensional sequential data and proposing a novel decoupling framework for static and dynamic representations.

Contribution

Introduces three GSN-based models for sequence learning, providing evidence that GSNs effectively learn complex sequential representations by decoupling static and dynamic components.

Findings

GSNs are viable for complex sequence representation

Models perform well on MNIST, bouncing balls, and motion capture data

Decoupling static and dynamic representations improves learning

Abstract

This thesis investigates unsupervised time series representation learning for sequence prediction problems, i.e. generating nice-looking input samples given a previous history, for high dimensional input sequences by decoupling the static input representation from the recurrent sequence representation. We introduce three models based on Generative Stochastic Networks (GSN) for unsupervised sequence learning and prediction. Experimental results for these three models are presented on pixels of sequential handwritten digit (MNIST) data, videos of low-resolution bouncing balls, and motion capture data. The main contribution of this thesis is to provide evidence that GSNs are a viable framework to learn useful representations of complex sequential input data, and to suggest a new framework for deep generative models to learn complex sequences by decoupling static input representations from dynamic time dependency representations.