Deep Variational Bayes Filters: Unsupervised Learning of State Space Models from Raw Data
This addresses the problem of learning state space models without supervision for applications like sequence prediction, though it appears incremental as it builds on existing variational inference methods.
The paper tackles unsupervised learning of latent Markovian state space models from raw data like image sequences, introducing Deep Variational Bayes Filters (DVBF) which uses variational inference to handle nonlinear dependencies. The result shows that backpropagation through transitions improves latent embedding information content and enables realistic long-term prediction.
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.