Latent Matters: Learning Deep State-Space Models
This work is significant for researchers and practitioners working with time-series data and dynamic systems, as it provides a more robust method for learning accurate underlying dynamics, which is an incremental improvement.
This paper addresses the issue that deep state-space models (DSSMs) often fail to learn underlying dynamics despite maximizing the evidence lower bound. The authors propose a constrained optimization framework and introduce the extended Kalman VAE (EKVAE), which significantly improves system identification and prediction accuracy compared to existing state-of-the-art DSSMs.
Deep state-space models (DSSMs) enable temporal predictions by learning the underlying dynamics of observed sequence data. They are often trained by maximising the evidence lower bound. However, as we show, this does not ensure the model actually learns the underlying dynamics. We therefore propose a constrained optimisation framework as a general approach for training DSSMs. Building upon this, we introduce the extended Kalman VAE (EKVAE), which combines amortised variational inference with classic Bayesian filtering/smoothing to model dynamics more accurately than RNN-based DSSMs. Our results show that the constrained optimisation framework significantly improves system identification and prediction accuracy on the example of established state-of-the-art DSSMs. The EKVAE outperforms previous models w.r.t. prediction accuracy, achieves remarkable results in identifying dynamical systems, and can furthermore successfully learn state-space representations where static and dynamic features are disentangled.