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

• DOI: 10.48550/arxiv.1605.06432
• Publication Date: 2016-01-01
• Authors: Maximilian Karl; Maximilian Soelch; Justin Bayer; Patrick van der S...
• Keywords: Backpropagation

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 Bayes, DVBF can overcome intractable inference distributions via variational inference. Thus, it handle highly nonlinear input data with temporal spatial dependencies such as image sequences without domain knowledge. Our experiments show that enabling backpropagation through transitions enforces assumptions significantly improves information content the embedding. This also enables realistic long-term prediction.