Linear Independent Component Analysis via Optimal Transport
The paper replaces ICA’s usual non-Gaussianity proxies with squared Wasserstein distance to a standard Gaussian.
Jha, Besserve, and Buchholz prove that this distance is maximized when a linear projection recovers an independent component. Their OT-ICA algorithm uses gradient-based optimization to find that projection. In simulations, it outperforms proxy-based ICA methods across different latent-variable distributions. The authors also test it on EEG artifact removal and econometric price discovery, arguing it works without distributional assumptions. ArXiv · AI/CL/LG's note
Jha, Besserve, and Buchholz prove that this distance is maximized when a linear projection recovers an independent component. Their OT-ICA algorithm uses gradient-based optimization to find that projection. In simulations, it outperforms proxy-based ICA methods across different latent-variable distributions. The authors also test it on EEG artifact removal and econometric price discovery, arguing it works without distributional assumptions. ArXiv · AI/CL/LG's note
score 4