A Local-Ascending-Global Learning Strategy for Brain-Computer Interface

Dongrui Gao, Haokai Zhang, Pengrui Li, Tian Tang, Shihong Liu, Zhihong Zhou, Shaofei Ying, Ye Zhu, Yongqing ZhangAAAI 2024 · DOI

A Local-Ascending-Global Learning Strategy for Brain-Computer Interface

Introduction

Brain-Computer Interfaces (BCIs) aim to decode human intent or cognitive states from electroencephalogram (EEG) signals. However, the brain's complex and dynamic nature makes this a significant challenge. Current Graph Neural Network (GNN) approaches often struggle to balance local regional nuances with global network interactions. Local methods focus on specific regions but miss large-scale functional connectivity, while global methods often rely on static measures, failing to capture task-specific, dynamic relationships. This paper introduces the Local-Ascending-Global (LAG) learning strategy, a hierarchical approach that bridges this gap by progressively fusing local brain region data into a global network through a self-adaptive "ascending" process.

Methodology

The LAG strategy is composed of three primary modules designed to mimic the hierarchical information processing of the human brain.

The LAG strategy architecture, featuring Local Feature Extraction, the SALK network, and the Graph Learning Module.
  1. Local Feature Extraction: This module processes individual functional brain regions using multi-kernel convolutions and a Bidirectional Gated Recurrent Unit (BiGRU). This combination effectively captures both intra-regional spatial features and the temporal dynamics of EEG signals.
  2. K-Level Self-Adaptive Ascending Network (SALK): The core innovation of the LAG strategy, SALK facilitates hierarchical fusion. It employs Singular Value Decomposition (SVD) to reduce data redundancy and calculates correlation scores to identify strong functional connections for specific tasks. The network then "ascends" through K-levels, merging relevant regions to uncover latent global patterns.
  3. Graph Learning Module: The resulting global representation is processed by a Dynamic Graph Convolutional Neural Network (DGCNN) and an attention mechanism to produce the final classification.

Experimental Evaluation

The model was validated on three diverse EEG datasets to assess its cross-subject generalizability:

  • SEED-VIG: Focused on vigilance and fatigue detection.
  • SEED-IV: Focused on four-class emotion recognition (neutral, sad, fear, happy).
  • BCI Competition IV 2a: A four-class motor imagery dataset (imagining movement of left hand, right hand, feet, or tongue).

Results and Discussion

The LAG strategy achieved state-of-the-art (SOTA) performance in independent-subject experiments. In the SEED-VIG dataset, it reached 85.15% accuracy, surpassing previous benchmarks like LGGNet. For emotion recognition (SEED-IV), it achieved 78.03% accuracy, a significant ~4% improvement over existing methods like RGNN. In the challenging BCI IV 2a motor imagery task, it reached 50.95% accuracy, remaining highly competitive in a notoriously difficult cross-subject environment.

A key strength of the LAG model is its interpretability. By visualizing the attention weights, the researchers demonstrated that the model focuses on biologically relevant areas for different tasks. For instance, in emotion recognition, the model prioritizes frontal and temporal lobes, whereas fatigue detection relies more on the occipital region.

Topographic maps showing the model's attention across different emotional states (neutral, sad, fear, happy) in the SEED-IV dataset.

The "ascending" fusion process proved critical; ablation studies indicated that building the global map from the bottom up preserves vital local features that are often lost in flat global models. SVD further enhanced this by filtering noise during the fusion process, ensuring only significant functional relationships were propagated.

Conclusion

The LAG strategy offers a robust, biologically-inspired framework for EEG decoding. While it demonstrates significant performance gains, future work will address the computational overhead of multi-level SVD and hierarchical fusion to enable real-time applications. Additionally, integrating more psychological and biological priors could further refine the functional mapping process, leading to even more precise intent decoding.

Summary updated

BibTeX
@inproceedings{gao2024local,
  author = {Dongrui Gao and Haokai Zhang and Pengrui Li and Tian Tang and Shihong Liu and Zhihong Zhou and Shaofei Ying and Ye Zhu and Yongqing Zhang},
  title = {{A Local-Ascending-Global Learning Strategy for Brain-Computer Interface}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i9.28867},
}
Figure