Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Efficient Solutions for Mitigating Initialization Bias in Unsupervised Self-Adaptive Auditory Attention Decoding

Yuanyuan Yao, Simon Geirnaert, Tinne Tuytelaars, Alexander Bertrand

TL;DR

This work tackles initialization bias in unsupervised self-adaptive auditory attention decoding (AAD) from EEG, which traditionally relies on calibration data. It builds a CCA/GEVD-based unsupervised framework and introduces three efficient variants to mitigate bias: a two-encoder approach, a soft-labeling method, and a sum-initialization strategy for the single-encoder. The results demonstrate that these alternatives can achieve performance comparable to a cross-validated unbiased baseline while maintaining a constant and low computational cost, with sum-initialization especially strong for small datasets and soft labeling becoming competitive as data grows. The proposed methods enable real-time or time-adaptive neuro-steered hearing devices, and the authors provide public code for implementation.

Abstract

Decoding the attended speaker in a multi-speaker environment from electroencephalography (EEG) has attracted growing interest in recent years, with neuro-steered hearing devices as a driver application. Current approaches typically rely on ground-truth labels of the attended speaker during training, necessitating calibration sessions for each user and each EEG set-up to achieve optimal performance. While unsupervised self-adaptive auditory attention decoding (AAD) for stimulus reconstruction has been developed to eliminate the need for labeled data, it suffers from an initialization bias that can compromise performance. Although an unbiased variant has been proposed to address this limitation, it introduces substantial computational complexity that scales with data size. This paper presents three computationally efficient alternatives that achieve comparable performance, but with a significantly lower and constant computational cost. The code for the proposed algorithms is available at https://github.com/YYao-42/Unsupervised_AAD.

Efficient Solutions for Mitigating Initialization Bias in Unsupervised Self-Adaptive Auditory Attention Decoding

TL;DR

This work tackles initialization bias in unsupervised self-adaptive auditory attention decoding (AAD) from EEG, which traditionally relies on calibration data. It builds a CCA/GEVD-based unsupervised framework and introduces three efficient variants to mitigate bias: a two-encoder approach, a soft-labeling method, and a sum-initialization strategy for the single-encoder. The results demonstrate that these alternatives can achieve performance comparable to a cross-validated unbiased baseline while maintaining a constant and low computational cost, with sum-initialization especially strong for small datasets and soft labeling becoming competitive as data grows. The proposed methods enable real-time or time-adaptive neuro-steered hearing devices, and the authors provide public code for implementation.

Abstract

Decoding the attended speaker in a multi-speaker environment from electroencephalography (EEG) has attracted growing interest in recent years, with neuro-steered hearing devices as a driver application. Current approaches typically rely on ground-truth labels of the attended speaker during training, necessitating calibration sessions for each user and each EEG set-up to achieve optimal performance. While unsupervised self-adaptive auditory attention decoding (AAD) for stimulus reconstruction has been developed to eliminate the need for labeled data, it suffers from an initialization bias that can compromise performance. Although an unbiased variant has been proposed to address this limitation, it introduces substantial computational complexity that scales with data size. This paper presents three computationally efficient alternatives that achieve comparable performance, but with a significantly lower and constant computational cost. The code for the proposed algorithms is available at https://github.com/YYao-42/Unsupervised_AAD.

Paper Structure

This paper contains 10 sections, 11 equations, 1 figure, 2 algorithms.

Table of Contents

  1. Introduction
  2. Methods
  3. Baseline: Single-Encoder Version
  4. Two-Encoder Version
  5. Soft Version
  6. Sum-Initialized Single-Encoder
  7. Experiments
  8. Dataset and Hyperparameters
  9. Results
  10. Conclusion

Figures (1)

  • Figure 1: Transductive accuracy, inductive accuracy, and normalized CPU time (w.r.t. baseline) across training set sizes. Dots and bars show mean and standard deviation across subjects and random seeds. Note: the supervised model is not shown in the transductive setting as this would imply using training labels.