Deep Variational Bayes Filters: Unsupervised Learning of State Space Models from Raw Data

TL;DR

Deep Variational Bayes Filters (DVBF) is a novel unsupervised method for learning latent Markovian state space models from raw, highly nonlinear data like image sequences, enabling better long-term predictions.

Contribution

DVBF introduces a variational inference approach that handles complex nonlinear data without domain knowledge, improving latent representations and long-term prediction capabilities.

Findings

Enables unsupervised learning of state space models from raw data.

Improves the information content of latent embeddings.

Enables realistic long-term predictions.

Abstract

We introduce Deep Variational Bayes Filters (DVBF), a new method for unsupervised learning and identification of latent Markovian state space models. Leveraging recent advances in Stochastic Gradient Variational Bayes, DVBF can overcome intractable inference distributions via variational inference. Thus, it can handle highly nonlinear input data with temporal and spatial dependencies such as image sequences without domain knowledge. Our experiments show that enabling backpropagation through transitions enforces state space assumptions and significantly improves information content of the latent embedding. This also enables realistic long-term prediction.