Robust Sparse Bayesian Learning Based on Minimum Error Entropy for Noisy High-Dimensional Brain Activity Decoding

Technical Summary: Robust Sparse Bayesian Learning Based on Minimum Error Entropy for Noisy High-Dimensional Brain Activity Decoding

Problem Statement
Brain activity decoding faces two primary challenges: the "high-dimensional, low-sample-size" problem and the presence of complex, non-Gaussian noise. Traditional machine learning methods, such as Ordinary Least Squares (OLS), suffer from overfitting in high-dimensional settings (e.g., fMRI voxels or EEG/ECoG channels) where the number of features exceeds the number of training samples. While Sparse Bayesian Learning (SBL) offers a framework to automatically prune irrelevant features via Automatic Relevance Determination (ARD), conventional SBL relies on likelihood functions based on idealized data distributions (e.g., Gaussian or Bernoulli). These assumptions often fail to capture the complex, non-Gaussian noise mixtures inherent in neural recordings (e.g., physiological artifacts, environmental noise), leading to performance deterioration in real-world scenarios. Existing robust methods, such as those based on Maximum Correntropy Criterion (MCC), often lack explicit sparsity control or require complex hyperparameter tuning, making them less suitable for high-dimensional decoding.

Methodology
The authors propose SBL-MEE, a robust Sparse Bayesian Learning framework that integrates the Minimum Error Entropy (MEE) learning criterion into the SBL skeleton. The core innovation lies in reformulating the Bayesian inference process under a generalized Bayesian paradigm (also known as the Gibbs posterior framework).

1. Generalized Posterior Formulation: Instead of relying on a standard likelihood function derived from a probabilistic density assumption, the method uses the MEE-induced loss ($L_{MEE}$) to drive the posterior update. The generalized posterior is defined as:
   $$\tilde{p}(w, a|t) \propto \exp(-L_{MEE}) p(w|a)p(a)$$
   where $w$ represents model parameters, $a$ represents relevance parameters (sparsity control), and $p(w|a)p(a)$ are the hierarchical Gaussian priors. This formulation balances empirical risk minimization (via MEE) with sparsity regularization (via priors).

2. Handling Regression and Classification:

   • Regression: The MEE objective is approximated using Quantized MEE (QMEE) to reduce computational complexity from $O(N^2)$ to $O(MN)$, where $M$ is the size of a quantization codebook.
   • Classification: To address the multimodal nature of classification errors (false negatives, true positives, false positives), the authors employ Restricted MEE (RMEE). This utilizes a fixed codebook $\{0, -1, 1\}$ to explicitly model the three-modal error distribution, ensuring the optimization targets the correct error structure.

3. Variational Inference and Optimization: Since the generalized posterior is intractable, the authors derive a learning algorithm using Variational Inference with mean-field factorization. They employ Laplacian approximation to approximate the non-Gaussian posterior of the weights as a Gaussian distribution. The algorithm iteratively updates the model parameters ($w$) using gradient-based methods (fixed-point for regression, half-quadratic for classification) and updates the relevance parameters ($a$) to enforce sparsity.

Key Contributions

1. Novel Framework: The proposal of SBL-MEE, a unified framework applicable to both regression and classification, which utilizes MEE as the optimization target for posterior updating within a generalized Bayesian setting.
2. Algorithmic Derivation: The derivation of a tractable learning algorithm using variational inference and Laplacian approximation to estimate model parameters for the MEE-based generalized posterior.
3. Comprehensive Evaluation: Evaluation on two real-world brain decoding tasks:
   • Regression: ECoG-based movement trajectory reconstruction (monkey data).
   • Classification: fMRI-based visual stimulus reconstruction (human data).
4. Interpretability Analysis: Demonstration that the proposed method yields decoder patterns that are not only statistically superior but also more physiologically plausible compared to existing methods.

Experimental Results
The SBL-MEE algorithm was compared against conventional SBL (Gaussian/Bernoulli likelihoods), the state-of-the-art SBCL (MCC-based), Ridge regression, and Lasso.

• Performance: In both regression (ECoG) and classification (fMRI) tasks, SBL-MEE achieved superior decoding performance, evidenced by higher correlation coefficients and lower Root Mean Squared Errors (RMSE) compared to all baselines. This superiority held even under challenging conditions with significantly reduced training data sizes (high-dimensional, low-sample-size scenarios).
• Physiological Interpretability:
  • ECoG (Motor): SBL-MEE identified spatial weights concentrated in motor-relevant cortical areas (e.g., Arcuate and Principal sulci), temporal weights focused on lags near movement onset, and frequency weights dominated by high-gamma bands. These patterns aligned closely with known neuroscientific principles of motor control, whereas other methods produced diffuse or less specific patterns.
  • fMRI (Visual): SBL-MEE produced retinotopic maps with strong diagonal structures (eccentricity matching) and weight distributions favoring the primary visual cortex (V1) and central visual fields. This demonstrated a higher degree of physiological plausibility compared to the uniform or overly spread patterns of Ridge, Lasso, and standard SBL.

Significance
The paper claims that SBL-MEE provides a powerful tool for improving brain activity decoding, particularly in noisy, high-dimensional settings typical of Brain-Computer Interface (BCI) and neuroscience research. By decoupling sparsity control (handled by the Bayesian prior) from noise robustness (handled by the MEE criterion), the method avoids the need for manual sparsity tuning while effectively handling non-Gaussian noise. The authors emphasize that while Minimum Error Entropy is not constructed from an arbitrary probabilistic distribution, it is highly effective for establishing noise-robust inference within the sparse Bayesian learning method, thereby promoting biomedical engineering applications.