Blind source separation using temporal predictability

Area of Science:

• Signal Processing
• Statistical Signal Analysis

Background:

• Linear mixtures of signals are common in various fields.
• Separating these mixtures into original source signals is a challenging problem.
• Existing methods like Independent Component Analysis (ICA) often rely on strong assumptions about source signal properties.

Purpose of the Study:

• To define and utilize a measure of temporal predictability for signal separation.
• To propose a novel method for recovering statistically independent source signals from their linear mixtures.
• To demonstrate the effectiveness of this method across diverse signal types and probability distributions.

Main Methods:

• Defining a novel measure of temporal predictability for signals.
• Developing an un-mixing matrix that maximizes temporal predictability for recovered signals.
• Solving the un-mixing matrix via a generalized eigenvalue problem with O(N^3) complexity.

Main Results:

• Demonstrated that temporal predictability of a mixture is less than or equal to its source signals.
• Successfully recovered source signals from linear mixtures with supergaussian, subgaussian, and gaussian probability density functions.
• Showcased the method's applicability to real-world mixtures, including voices and music.

Conclusions:

• Temporal predictability offers a robust criterion for separating linear signal mixtures.
• This method provides an alternative to ICA with fewer assumptions on source signal distributions.
• The approach is computationally feasible and effective for complex signal separation tasks.