A Discriminative and Robust Feature Learning Approach for EEG-Based Motor Imagery Decoding (Student Abstract)

Xiuyu Huang, Nan Zhou, Kup-Sze ChoiAAAI 2022 · DOI

Introduction

Decoding Motor Imagery (MI)—the process of identifying specific movements a person is imagining based on their brain waves—is a fundamental challenge in Brain-Computer Interface (BCI) research. While deep learning models like Convolutional Neural Networks (CNNs) have significantly improved decoding performance, Electroencephalogram (EEG) data remains notoriously difficult to process. It is characterized by a low signal-to-noise ratio and is frequently contaminated by non-Gaussian noise and outliers, such as eye blinks or electrical interference.

To improve classification, researchers often employ "Center Loss," a technique that encourages the network to minimize the distance between features of the same class and their respective class centers. This results in tighter, more discriminative clusters. However, standard Center Loss relies on the $L_2$ norm (squared Euclidean distance). Because the $L_2$ penalty grows quadratically with distance, even a single outlier far from the center can disproportionately influence the gradient, "pulling" the entire class center toward the noise and degrading performance.

Correntropy based Center Loss (CCL)

To address the sensitivity of standard Center Loss to outliers, this paper proposes Correntropy based Center Loss (CCL). The core innovation lies in replacing the $L_2$ distance with a measure based on the Maximum Correntropy Criterion (MCC).

Instead of penalizing the raw squared distance $||x_i - c_{y_i}||^2$, the CCL function utilizes a Gaussian kernel: $$L_{CCL} = \sum_{i=1}^{N} (1 - \exp(-\frac{||x_i - c_{y_i}||^2}{2\sigma^2}))$$

The Robustness Mechanism

The use of the exponential kernel introduces a "saturation effect." As a data point moves further from its class center, the loss value eventually levels off. This means that extreme outliers have a limited influence on the gradient, allowing the model to effectively ignore them rather than attempting to accommodate them by shifting the cluster center. The kernel size parameter, $\sigma$, acts as a controllable dial for robustness, determining the range within which feature points are considered relevant for clustering.

Visualization of feature distributions on a toy MNIST dataset. (c) Standard Center Loss is heavily distorted by noise, whereas (d) the proposed CCL method maintains tight, well-separated clusters by isolating noise in the center of the feature space.

Optimization and Joint Supervision

The authors employ a two-step optimization strategy to train the network. Deep features are updated via standard Gradient Descent, while class centers are updated using a "Half Quadratic" strategy. To ensure both intra-class compactness and inter-class separability, the final training objective is a joint supervision signal combining standard Softmax loss with the proposed CCL.

Experimental Evaluation

The effectiveness of CCL was evaluated across three distinct scenarios: a toy MNIST example for visualization and two competitive BCI datasets.

  1. MNIST Toy Example: Used to demonstrate how the loss handles artificial noise (treating one digit class as noise distributed across others). Visualizations confirm that while standard Center Loss clusters are shifted by noise, CCL clusters remain stable.
  2. BCI Competition IV-2a: A four-class motor imagery dataset (left hand, right hand, feet, tongue).
  3. BCI Competition IV-2b: A two-class motor imagery dataset (left hand vs. right hand).
Feature distribution for the BCI Competition IV-2a dataset using the CCL method. The visualization shows clear separation between the four motor imagery classes and the noise points.

Key Results

The CCL-enhanced model demonstrated superior performance compared to state-of-the-art baselines:

  • BCI IV-2a: Achieved an accuracy of 79.07%, significantly outperforming the standard EEGNet baseline (74.50%).
  • BCI IV-2b: Achieved 85.29% accuracy, surpassing the ConNet baseline (79.37%).
  • Noise Resistance: The model showed remarkable resilience to 50Hz power line interference, a common artifact in EEG recording, with the feature distribution remaining largely unaffected.

Conclusion and Future Directions

The Correntropy based Center Loss provides a robust alternative to traditional distance-based clustering for noisy signal processing. By leveraging the saturation properties of correntropy, the model achieves high discriminative power without being compromised by the non-Gaussian noise inherent in EEG data.

As a student abstract, this work opens several avenues for future research. Key areas include exploring the sensitivity of the model to hyperparameters like the kernel size $\sigma$, evaluating the computational overhead of the two-step update process, and testing the framework's performance in real-time BCI applications.

Summary updated

BibTeX
@inproceedings{huang2022discriminative,
  author = {Xiuyu Huang and Nan Zhou and Kup-Sze Choi},
  title = {{A Discriminative and Robust Feature Learning Approach for EEG-Based Motor Imagery Decoding (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21622},
}
Figure