[Paper Review] Hidden Markov Nonlinear ICA: Unsupervised Learning from Nonstationary Time Series

Hermanni Hälvä, Aapo Hyvärinen|arXiv (Cornell University)|Jun 22, 2020

Blind Source Separation Techniques25 references17 citations

TL;DR

This paper proposes Hidden Markov Nonlinear ICA, a fully unsupervised framework that replaces manual segmentation of nonstationary time series with a latent Hidden Markov Model (HMM) to model time-varying data distributions. By integrating HMMs into nonlinear ICA, the method achieves identifiability, enables end-to-end learning via EM, and outperforms prior methods like TCL in recovering independent components and modeling temporal dynamics on synthetic data.

ABSTRACT

Recent advances in nonlinear Independent Component Analysis (ICA) provide a principled framework for unsupervised feature learning and disentanglement. The central idea in such works is that the latent components are assumed to be independent conditional on some observed auxiliary variables, such as the time-segment index. This requires manual segmentation of data into non-stationary segments which is computationally expensive, inaccurate and often impossible. These models are thus not fully unsupervised. We remedy these limitations by combining nonlinear ICA with a Hidden Markov Model, resulting in a model where a latent state acts in place of the observed segment-index. We prove identifiability of the proposed model for a general mixing nonlinearity, such as a neural network. We also show how maximum likelihood estimation of the model can be done using the expectation-maximization algorithm. Thus, we achieve a new nonlinear ICA framework which is unsupervised, more efficient, as well as able to model underlying temporal dynamics.

Motivation & Objective

• To address the limitation of existing nonlinear ICA methods that require observed auxiliary variables like segment indices for nonstationary data.
• To develop a fully unsupervised framework that infers latent temporal states from data without manual segmentation.
• To enable identifiability of independent components when the conditioning variable is unobserved and latent.
• To model repeating nonstationary dynamics efficiently using a small number of hidden states.
• To provide a unified framework for unsupervised representation learning and temporal dynamics inference in time series.

Proposed method

• Integrates a Hidden Markov Model (HMM) as a latent conditioning variable in the nonlinear ICA framework, replacing observed segment indices.
• Uses the HMM to model time-varying data distributions via discrete latent states with Markov transition dynamics.
• Proves identifiability of the model under general mixing nonlinearities, including deep neural networks, by extending theoretical results from Hyvärinen et al. (2019).
• Employs the Expectation-Maximization (EM) algorithm for maximum likelihood estimation, with M-step using Adam optimizer on mini-batches of sub-sequences.
• Parameterizes the inverse mixing function using a deep neural network (MLP) with leaky ReLU activations and weight sharing across layers.
• Applies the Viterbi algorithm to infer the most likely sequence of latent states after model training.

Experimental results

Research questions

• RQ1 Can a nonlinear ICA model be made fully unsupervised by replacing observed segment indices with a latent HMM?
• RQ2 Is identifiability of the independent components preserved when the conditioning variable is unobserved and inferred via HMM?
• RQ3 Can the proposed model outperform existing methods like TCL in recovering true independent components from nonstationary time series?
• RQ4 To what extent can the model learn and reconstruct the underlying temporal dynamics of latent states?
• RQ5 Does the use of a latent HMM improve estimation efficiency and accuracy compared to fixed, manually segmented intervals?

Key findings

• The proposed Hidden Markov Nonlinear ICA model achieves superior performance in recovering true independent components compared to Time-Contrastive Learning (TCL), especially under higher nonlinearity (L=4), where it significantly outperforms TCL across all tested configurations.
• For N=5 and C=11, the model achieves near-perfect accuracy (close to 100%) in latent state classification under linear ICA, 80% accuracy for L=2, and 68% for L=4, indicating strong temporal dynamics recovery.
• The model successfully identifies and clusters latent states in a way that reflects the true underlying dynamics, even when the number of states is small and repeating over time.
• The use of stochastic mini-batches in the EM algorithm leads to faster convergence, demonstrating practical scalability.
• The identifiability of the model is proven under general mixing nonlinearities, including deep neural networks, establishing theoretical soundness.
• The framework enables joint learning of disentangled features and temporal dynamics, offering a more principled and efficient approach to unsupervised representation learning on nonstationary data.