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

**Authors:** Maximilian Karl, Maximilian Soelch, Justin Bayer, Patrick van der Smagt  
**arXiv:** [1605.06432](https://arxiv.org/abs/1605.06432)  
**Date:** 2016-05-20  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/1605.06432v3/

## Key points

- The paper introduces a new method called Deep Variational Bayes Filters (DVBF) for unsupervised learning and identification of latent Markovian state space models.
- DVBF leverages Stochastic Gradient Variational Bayes to overcome intractable inference distributions through variational inference.
- DVBF can handle highly nonlinear input data with temporal and spatial dependencies, such as image sequences, without requiring domain knowledge.
- Enabling backpropagation through transitions improves the information content of the latent embedding and enables realistic long-term prediction.
- DVBF is a powerful tool for unsupervised learning and identification of latent Markovian state space models.
- It is particularly useful for analyzing image sequences.
- Experimental results demonstrate the effectiveness of DVBF in improving the information content of the latent embedding and enabling accurate long-term predictions.

## Summary

The paper "Deep Variational Bayes Filters: Unsupervised Learning of State Space Models from Raw Data" introduces a new method called Deep Variational Bayes Filters (DVBF) for unsupervised learning and identification of latent Markovian state space models. DVBF leverages recent advances in Stochastic Gradient Variational Bayes to overcome intractable inference distributions through variational inference. This allows DVBF to handle highly nonlinear input data with temporal and spatial dependencies, such as image sequences, without requiring domain knowledge. The authors conducted experiments to evaluate the performance of DVBF and found that enabling backpropagation through transitions enforces state space assumptions and significantly improves the information content of the latent embedding. This improvement enables realistic long-term prediction. Overall, DVBF provides a powerful tool for unsupervised learning and identification of latent Markovian state space models. Its ability to handle complex data with temporal and spatial dependencies makes it particularly useful for tasks such as analyzing image sequences. The experimental results demonstrate the effectiveness of DVBF in improving the information content of the latent embedding and enabling accurate long-term predictions.