Time Series Source Separation with Slow Flows

TL;DR

This paper integrates slow feature analysis into flow-based models to improve the identifiability of time series source separation, leveraging invertible neural models for better decomposition.

Contribution

It introduces a novel approach combining SFA with flow-based models, enhancing source separation and identifiability in time series analysis.

Findings

SFA fits naturally into flow-based models.

The approach improves source separation in time series.

Identifiability of decomposed sources is achieved.

Abstract

In this paper, we show that slow feature analysis (SFA), a common time series decomposition method, naturally fits into the flow-based models (FBM) framework, a type of invertible neural latent variable models. Building upon recent advances on blind source separation, we show that such a fit makes the time series decomposition identifiable.