Unsupervised Hidden State Estimation and Blind Source Separation Using Auto-Encoder RNN Filter

EasyChair Preprint 9915

Abstract

This work proposes a deep learning estimator for unsupervised nonlinear hidden state estimation formulating the problem as blind source separation (BSS). The model is composed of an auto encoder base d RNN to estimate the hidden state. The model is extended to the blind source separation using local losses to de-correlate the hidden signals. The problem is formulated such that the number of sources can be determined by varying the dimension of the hidden state signal. The solution is demonstrated on a number of simulations. The simulation shows that the model is suitable for hidden state extraction. We find that the model will extract the hidden signals correctly when the correct dimensionality is selected , otherwise repeated hidden signals occur. Similarly, when applied to BSS, the model successfully separated multiple sources. The model retains many of the limitations of BSS, such as being able to recover the component signals but not amplitude. The use of an auto-encoder limits the model to cases of over specified problems, where more sensors than hidden states are present, making it well suited for domains with multiple redundant sensors (drones, self-driving cars, etc.) The filter provides much functionality by de-noising sensor signals, decoding sensor signals to either (1) a lower dimensional latent space performing (non-linear PCA) or (2) separates source signals (non-linear ICA) and forecasts predictions.

Keyphrases: Auto-encoders, Hidden State Estimation, Recurrent Neural Networks, blind source separation., filtering, unsupervised learning

@booklet{EasyChair:9915,
  author    = {Clint Steed and Kim Namhun},
  title     = {Unsupervised Hidden State Estimation and Blind Source Separation Using Auto-Encoder RNN Filter},
  howpublished = {EasyChair Preprint 9915},
  year      = {EasyChair, 2023}}