Reproduce “Temporally Disentangled Representation Learning under Unknown Nonstationarity” on OpenResearch

About this paper

Xiangchen Song, Weiran Yao, Yewen Fan, Xinshuai Dong, Guangyi Chen, Juan Carlos Niebles, Eric Xing, Kun Zhang

This research from Carnegie Mellon University and Salesforce Research establishes the first theoretical identifiability results for learning disentangled representations from sequential data where the underlying causal factors exhibit time-delayed dependencies and the nonstationarity is unobserved. The proposed NCTRL framework robustly recovers latent causal variables and their relationships, achieving superior performance on synthetic data and demonstrating efficacy on real-world video datasets like Modified CartPole and MoSeq.

Key insights

• Latent causal processes within sequential data are identifiable even under unobserved nonstationarity and with time-delayed causal dependencies, provided certain 'linear independence' conditions on the nonstationary domain-specific dynamics hold.
• Modeling unobserved nonstationary domain shifts as a first-order Markov Chain enables their identifiability and, consequently, the identifiability of the underlying latent causal factors.
• Accurately recovering disentangled latent representations in nonstationary sequential data requires simultaneously accounting for both the unobserved domain shifts and the time-delayed causal relationships among the latent variables themselves.

Summary by alphaXiv.