Reproduce “Unsupervised Feature Extraction by Time-Contrastive Learning and Nonlinear ICA” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Aapo Hyvarinen, Hiroshi Morioka
TCL trains features by labeling observations with their time-window indices and predicting those labels. Under assumptions on independent sources whose distributions change across windows, it can recover transformed sources after a nonlinear mixture, up to an invertible linear transformation. In simulations, TCL tracked transformed sources better than compared methods in nonlinear mixtures; this does not establish recovery of real-world sources.
Key insights
- Many real-world time series exhibit natural nonstationarity, with statistical properties changing over time.
- By explicitly modeling and leveraging temporal structure and training a network to discriminate between time segments, nonlinear source separation can become theoretically identifiable.
- The features learned by Time-Contrastive Learning (TCL) are proven to be identifiable, representing nonlinear functions of the original independent sources up to a linear transformation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.