Research map / Probabilistic and scientific modeling / Dynamical systems
A Local-Ascending-Global Learning Strategy for Brain-Computer Interface
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LAG, a local-to-global EEG learning strategy with a K-level self-adaptive ascending network (SALK) that progressively fuses brain regions and feeds a DGCNN classifier, evaluated on fatigue, emotion, and motor imagery datasets.
Problem
Decoding EEG for BCI cognitive tasks (fatigue, emotion, motor imagery) requires capturing both local connectivity within brain regions and adaptive higher-level connectivity among regions, which existing fixed local or global graph methods do not model.
Approach
Builds local features per brain region (ghost module, frequency/spatial attention, BiGRU), then iteratively fuses region pairs via SVD-based fusion and a correlation score that selects the most strongly connected joint region at each level, ascending from local to global features before a DGCNN graph classifier.
Key results
- LAG reaches 85.15% ACC and 87.56% F1 on SEED-VIG, outperforming LGGNet (83.89 / 85.48).
- LAG reaches 78.03% ACC on SEED-IV with 07.38 standard deviation, the lowest std among compared methods.
- On BCI IV 2a LAG achieves 50.95% accuracy, trailing R-SPD (51.2%) by 0.25%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply LAG instead of SFTNet | SEED-VIG recognition · SEED-VIG | ACC: 83.18 -> 85.15 | Helped |
| apply LAG instead of SFTNet | SEED-VIG recognition · SEED-VIG | F1-score: 86.97 -> 87.56 | Helped |
| apply LAG instead of FGI | SEED-IV emotion recognition · SEED-IV | ACC: 77.53 -> 78.03 | Helped |
| apply LAG instead of R-SPD | BCI C IV 2a motor imagery · BCI C IV 2a | accuracy: 51.2 -> 50.95 | Hurt |
| use adaptive ascending (SALK) instead of Local fusion (ablation) | SEED-VIG fatigue recognition, 6 brain regions · SEED-VIG | ACC: 81.93 -> 85.15 | Helped |
| use adaptive ascending (SALK) instead of Global fusion (ablation) | SEED-VIG fatigue recognition, 6 brain regions · SEED-VIG | ACC: 82.99 -> 85.15 | Helped |
| use adaptive ascending (SALK) instead of Local fusion (ablation) | SEED-VIG fatigue recognition, 6 brain regions · SEED-VIG | F1-score: 84.85 -> 87.56 | Helped |
| use adaptive ascending (SALK) instead of Global fusion (ablation) | SEED-VIG fatigue recognition, 6 brain regions · SEED-VIG | F1-score: 84.64 -> 87.56 | Helped |
| increase SALK brain regions from 1 to 6 (ablation) | SEED-VIG fatigue recognition · SEED-VIG | ACC: 78.38 -> 85.15 | Helped |
| increase SALK brain regions from 1 to 6 (ablation) | SEED-VIG fatigue recognition · SEED-VIG | F1-score: 79.24 -> 87.56 | Helped |
Surprising findings
- On BCI IV 2a, LAG is marginally below R-SPD, so it is not best on all datasets.
Limitations
- Gains are modest (about 3-4% on average) and LAG trails R-SPD on BCI IV 2a.
- Evaluation is limited to independent-subject (cross-subject) setting on three datasets.
- Baselines for non-released datasets were reproduced by the authors, which may affect fairness.
Experimental setup
Datasets: SEED-VIG SEED-IV BCI Competition IV 2a
Benchmarks: SEED-VIG SEED-IV BCI Competition IV 2a
Metrics: accuracy F1-score standard deviation
Baselines: CSF-GTNet Interpretable CNN SFTNet E-5-D-BCDRNet LGGNet RGNN JTSR FGI RGCBNet PGCN OVR-FBCSP R-SPD
Builds on: DGCNN SparseDGCNN LGGNet Ghost module Relation-aware global attention
Closest papers by meaning
- CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding (NeurIPS 2025 · similarity 0.87)
- MindSimulator: Exploring Brain Concept Localization via Synthetic fMRI (ICLR 2025 · similarity 0.86)
- Beyond Pairwise Connections: Extracting High-Order Functional Brain Network Structures under Global Constraints (NeurIPS 2025 · similarity 0.86)
- Functional Connectomes of Neural Networks (AAAI 2025 · similarity 0.85)
- Learning Time-Varying Multi-Region Brain Communications via Scalable Markovian Gaussian Processes (ICML 2025 · similarity 0.85)
- Brain-Like Processing Pathways Form in Models With Heterogeneous Experts (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.