Robust Sparse Blind Source Separation

Cecile Chenot, Jérôme Bobin, Jérémy Rapin

TL;DR

The paper introduces rGMCA (robust Generalized Morphological Component Analysis), a new algorithm for Blind Source Separation (BSS) that remains accurate even when data is corrupted by large, unpredictable outliers. By jointly estimating the sources, the mixing matrix, and the outliers—and using a clever weighting scheme to "ignore" corrupted samples—rGMCA significantly outperforms standard techniques in noisy, real-world scenarios.

What Problem Does This Paper Address?

Blind Source Separation (BSS) is the "cocktail party problem" of data science: given several mixed signals (like audio from multiple microphones), how do you recover the original individual sources? While standard methods like Independent Component Analysis (ICA) or sparse BSS work well in ideal conditions, they are notoriously fragile.

In fields like astrophysics or medical imaging, data is often plagued by outliers—spurious spikes, sensor glitches, or impulsive noise—that don't fit a standard Gaussian distribution. Most BSS tools try to "average out" noise, but a single massive outlier can completely tilt the estimation of the mixing matrix, leading to poor source recovery. Existing solutions either try to scrub outliers beforehand (which is risky) or use specialized math that only works for non-negative data (like NMF). There was a clear need for a robust, general-purpose sparse BSS method.

Methods

The authors build upon an existing algorithm called GMCA, which uses the "sparsity" of signals (the idea that most signals are mostly zero in some domain) to separate them.

The core innovation is the rGMCA algorithm, which introduces two main layers of defense against outliers:

  1. Joint Estimation: Instead of cleaning the data first, rGMCA treats outliers as a third component to be estimated alongside the sources and the mixing matrix ($X = AS + O + N$).
  2. Iterative Weighting Scheme: This is the "secret sauce." During the step where the algorithm estimates the mixing matrix ($A$), it applies a weighting matrix ($W$). This matrix assigns low weights to data samples that are likely contaminated by outliers. Essentially, the algorithm learns which parts of the data are "untrustworthy" and ignores them when figuring out how the sources are mixed.

The algorithm uses a "decreasing threshold" strategy, starting by looking only at the strongest, most distinct parts of the signal and gradually refining the estimate as the noise and outliers are stripped away.

Datasets

The researchers evaluated rGMCA using two primary setups:

  • Simulated Synthetic Data: They created mixtures of 8 sources across 16 observations using Bernoulli-Gaussian distributions. They intentionally added "corrupted columns" and random spikes to test robustness against different outlier patterns.
  • NMR Spectroscopy Data: They simulated Nuclear Magnetic Resonance (NMR) spectra for four molecules (menthone, folic acid, ascorbic acid, and myo-inositol). This provided a realistic test case for biomedical engineering, where identifying chemical signatures in a mixture is critical.

Key Insights and Ideas

  • Outliers Break the MDP: Sparse BSS relies on the Morphological Diversity Property (MDP)—the idea that different sources don't have their biggest peaks at the same time. Large outliers break this because they become the dominant "peaks," tricking standard algorithms into thinking the outlier is a source.
  • Weighting > Filtering: Simply trying to subtract outliers (the "Naive" approach) fails when outliers are large. By using a weighting scheme to downweight corrupted samples during the mixing matrix update, rGMCA prevents the "pollution" of the spatial coordinate system.
  • Robustness to Sample Size: Unlike methods based on Principal Component Pursuit (PCP), which require many more observations than sources ($m \gg n$), rGMCA remains effective even when the number of observations is close to the number of sources.

Key Results

  • Superior Accuracy: In simulations, rGMCA maintained a low error rate ($\Delta_A$) even when the amplitude of outliers was $10,000$ times larger than the signal. Standard GMCA and $\beta$-divergence methods failed as soon as outlier amplitude increased.
  • High Breakdown Point: rGMCA could handle up to 5-10% total data corruption while maintaining a near 100% "probability of success" in recovering the mixing matrix.
  • Real-world Application: In the NMR tests, rGMCA was the only algorithm able to provide a correct estimate of the mixing matrix when the number of observations was low ($m \ge 5$ for 4 sources), whereas other robust methods required significantly more data.

Limitations and Future Work

While rGMCA is excellent at estimating the mixing matrix (the "how" of the mixture), the authors note that it still struggles to perfectly separate the original outliers from the estimated sources in the final output (as seen in the NMR results). The recovered sources can still contain some residual artifacts from the outliers.

Future Work: The authors plan to generalize this approach to work in "transformed domains" (like wavelets), which would allow the algorithm to handle sources that aren't sparse in their raw form but become sparse when mathematically transformed.

Summary updated

Abstract

Blind source separation is a widely used technique to analyze multichannel data. In many real-world applications, its results can be significantly hampered by the presence of unknown outliers. In this paper, a novel algorithm coined rGMCA (robust Generalized Morphological Component Analysis) is introduced to retrieve sparse sources in the presence of outliers. It explicitly estimates the sources, the mixing matrix, and the outliers. It also takes advantage of the estimation of the outliers to further implement a weighting scheme, which provides a highly robust separation procedure. Numerical experiments demonstrate the efficiency of rGMCA to estimate the mixing matrix in comparison with standard BSS techniques.
BibTeX
@article{chenot2015robust,
  author = {Cecile Chenot and Jérôme Bobin and Jérémy Rapin},
  title = {{Robust Sparse Blind Source Separation}},
  journal = {IEEE Signal Processing Letters},
  year = {2015},
  url = {https://openreview.net/forum?id=ycfNKfhzAd},
}

Loading related content...

Figure