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

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

Problem Statement
Auditory Attention Decoding (AAD) aims to identify the speaker a user is attending to in a multi-speaker environment using electroencephalography (EEG) signals, a capability critical for neuro-steered hearing aids. Current state-of-the-art methods typically require supervised training with ground-truth labels of the attended speaker, necessitating tedious calibration sessions for every user and EEG setup. To eliminate this dependency, unsupervised self-adaptive AAD based on stimulus reconstruction has been developed. This approach initializes a decoder with random attention labels and iteratively refines the model using newly predicted labels.

However, this unsupervised self-adaptive process suffers from initialization bias. Because the model trains and tests on the same data across iterations, it tends to reinforce initial (potentially incorrect) predictions, causing the system to get stuck in a suboptimal state. While a previous solution by Heintz et al. [8] addressed this by implementing leave-one-out cross-validation within each iteration, this approach introduces substantial computational complexity that scales linearly with the number of data segments ($K$), making it impractical for real-time or time-adaptive applications.

Methodology
The paper proposes three computationally efficient alternatives based on Canonical Correlation Analysis (CCA) that mitigate initialization bias without the heavy computational overhead of cross-validation. The methods operate on a 2-speaker scenario (generalizable to $N$ speakers) where EEG signals ($X$) and speech features ($S_1, S_2$) are processed in segments.

1. Two-Encoder Version:
   Unlike the baseline single-encoder model which only optimizes for the attended speaker's features ($S_a$), this variant introduces a second encoder for the unattended speaker's features ($S_u$). Both encoders share a single EEG decoder ($W_x$). The objective is to jointly maximize the correlation between the EEG and the features of both speakers. The hypothesis is that by forcing the decoder to account for responses to both speakers, the model becomes less susceptible to being "attracted" to wrongly labeled attended speech, thereby increasing the robustness to label errors and the ability to recover from initial mistakes.

2. Soft Version:
   Instead of making hard binary assignments (attending to Speaker 1 or 2) for each segment, this method assigns probabilistic weights ($p_{1k}, p_{2k}$) based on prediction uncertainty. The statistics used for CCA ($R_{xa}$ and $R_{aa}$) are computed using a weighted sum of the speaker features ($p_{1k}S_{1k} + p_{2k}S_{2k}$). The probabilities are estimated using a subject-specific, unsupervised approach that models the distribution of canonical correlations for attended and unattended speakers as two Gaussian distributions. This provides a principled middle ground, maintaining the optimal single-encoder structure while adaptively incorporating information from both speakers when the model is uncertain.

3. Sum-Initialized Single-Encoder:
   This approach addresses the bias at the very first step. Rather than initializing with random labels, the model is trained in the first iteration using a composite signal formed by the sum of both speakers' features. This corresponds to setting the initial probabilities to $p_{1k} = p_{2k} = 0.5$. The hypothesis is that this provides an informative starting point that captures neural responses common to both speakers, preventing the self-sustaining label bias that arises from arbitrary random initialization.

Key Contributions

• Algorithmic Efficiency: The paper presents three methods that achieve performance comparable to the cross-validated baseline [8] but with significantly lower and constant computational costs.
• Scalability: Unlike the cross-validated method, whose computational time scales linearly with the dataset size (requiring $K$ model trainings per iteration), the proposed methods require only a single model training per iteration.
• Bias Mitigation: The proposed variants successfully break the cycle of initialization bias inherent in standard self-adaptive AAD, allowing the model to converge to higher-quality solutions without expensive inner-loop cross-validation.

Experimental Results
The methods were evaluated on a public dataset containing 64-channel EEG recordings from 16 subjects attending to one of two competing speakers. Performance was measured in both transductive (predicting on training data) and inductive (generalizing to unseen data) settings.

• Accuracy:
  • The Sum-Initialized Single-Encoder consistently outperformed the Two-Encoder method and was particularly effective with limited training data (5–15 minutes).
  • The Soft Version underperformed on small datasets but approached the accuracy of the cross-validated variant as the training set size increased.
  • In general, removing initialization bias yielded more pronounced improvements in transductive decoding.
• Computational Cost:
  • The cross-validated baseline's CPU time scaled linearly, reaching approximately 30 times the baseline cost for 45-minute training sets.
  • The Two-Encoder and Soft methods maintained a constant normalized CPU time of approximately 1.5× the baseline, regardless of data size.
  • The Sum-Initialized method matched the baseline's computational cost (1.0×).

Significance and Claims
The authors claim that these solutions make unsupervised self-adaptive AAD significantly more viable for real-time or time-adaptive implementations [9]. By eliminating the need for expensive cross-validation while maintaining robustness against initialization bias, the proposed methods offer a practical pathway for deploying neuro-steered hearing devices without the burden of frequent, labeled calibration sessions. The paper emphasizes that for smaller datasets, the Sum-Initialized approach is the top performer with minimal computational overhead, while the Soft-label method becomes competitive for larger datasets, offering a low-cost alternative to the cross-validated gold standard.