Ungrounded independent non-negative factor analysis

Bhiksha Raj, Kevin W. Wilson, Alexander Krueger, Reinhold Haeb-UmbachInterspeech 2010 · DOI

Abstract

This paper introduces Ungrounded Independent Non-Negative Factor Analysis (U-INFA), a novel algorithm for decomposing non-negative mixtures into independent source components. Unlike existing non-negative Independent Component Analysis (ICA) methods that require signals to be "grounded" (periodically reaching zero), U-INFA handles "ungrounded" signals with constant DC offsets. By integrating a "shift-to-origin" preprocessing step with a decorrelation-regularized Non-Negative Matrix Factorization (NMF) framework, the proposed method demonstrates superior robustness and accuracy in noisy, real-world source separation tasks.

Introduction

In signal processing, decomposing a complex mixture—such as an audio spectrogram—into its original independent sources is a fundamental challenge. Standard tools like Independent Component Analysis (ICA) and Non-Negative Matrix Factorization (NMF) possess inherent limitations. ICA often produces physically impossible negative values for non-negative energy signals, while NMF lacks guarantees regarding the statistical independence of its components.

Furthermore, existing non-negative ICA algorithms typically rely on the "groundedness" assumption, where sources must frequently hit zero. In real-world environments, signals often contain persistent background levels or DC offsets, causing these traditional algorithms to fail. U-INFA addresses this "ungrounded" problem by enabling the recovery of independent components even when sources never reach zero.

Methodology

The U-INFA framework consists of three primary stages:

  1. Preprocessing (Artificial Grounding): The authors provide a mathematical proof that shifting ungrounded data toward the origin by subtracting the observed minimum value (infimum) allows the data to be treated as grounded.
  2. Core INFA Algorithm: A specialized NMF variant is applied, incorporating a decorrelation regularization term. This constraint ensures that the recovered source signals (rows of matrix $H$) are statistically independent. The algorithm directly estimates the mixing matrix ($W$), enhancing stability in noisy or overdetermined conditions.
  3. Postprocessing: The initial shift is reversed to recover the original ungrounded sources with their respective offsets.

Experimental Setup

The algorithm was evaluated using synthetic datasets designed to simulate various failure points:

  • Sources: 3-dimensional signals drawn from uniform distributions.
  • Offsets: A parameter $q$ (offset) was varied from 0 (grounded) to 2.0 (heavily ungrounded).
  • Mixing Scenarios: Both "square" mixing ($3 \times 3$) and "overdetermined" mixing ($6 \times 3$) were tested.
  • Noise: IID Gaussian noise was added to maintain a 19 dB input Signal-to-Noise Ratio (SNR).

Results and Discussion

U-INFA was compared against standard NMF, FastICA, and Plumbley’s Non-negative ICA.

Robustness to Offsets

In the square mixing scenario, the performance of standard non-negative ICA methods declined sharply as the signal offset increased. In contrast, U-INFA maintained a consistently high Output Signal-to-Interferer Ratio (SIR), demonstrating its effectiveness in handling ungrounded data.

Output SIR vs. Offset for 3x3 mixing.

Performance in Overdetermined Cases

In the $6 \times 3$ mixing scenario, where more observations are available than sources, U-INFA significantly outperformed all competitors. It achieved an average Output SIR between 14-16 dB, whereas the closest ungrounded competitor, U-Nneg ICA, reached only 10-12 dB.

Output SIR vs. Offset for 6x3 mixing.

The results indicate that U-INFA is not only more accurate on average but also more reliable in its "worst-case" performance (indicated by dashed lines in the figures), ensuring the recovery of all sources even in challenging conditions.

Conclusion

U-INFA provides a robust solution for independent component extraction from non-negative, ungrounded data. By leveraging the mixing matrix and a decorrelation proxy for independence, it overcomes the limitations of traditional ICA and NMF. While the algorithm is subject to local minima common in NMF-based approaches, its performance in noisy and overdetermined environments is a significant advancement. Future work will explore its application in real-world acoustic environments, such as speech denoising in the presence of non-stationary background hums.

Summary updated

BibTeX
@inproceedings{raj2010ungrounded,
  author = {Bhiksha Raj and Kevin W. Wilson and Alexander Krueger and Reinhold Haeb-Umbach},
  title = {{Ungrounded independent non-negative factor analysis}},
  booktitle = {Interspeech},
  year = {2010},
  doi = {10.21437/interspeech.2010-123},
}
Figure