Subject-class joint gradient alignment based domain generalization framework in unseen subject motor imagery classification

Domain generalization (DG) based on meta-learning (ML) can effectively enhance the model’s generalization capability for unknown subjects in motor imagery-based brain-computer interface (MI-BCI). However, the current methods merely achieve balance across different subjects while neglecting inter-cla...

Full description

Saved in:
Bibliographic Details
Published in:Journal of physics. Conference series Vol. 3079; no. 1; pp. 12083 - 12089
Main Authors: Liu, Hanlin, Li, Zhi, Li, Mingai
Format: Journal Article
Language:English
Published: Bristol IOP Publishing 01.08.2025
Subjects:
Alignment
Boundaries
Categories
Classification
Human-computer interface
Imagery
Test sets
ISSN:1742-6588, 1742-6596
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Domain generalization (DG) based on meta-learning (ML) can effectively enhance the model’s generalization capability for unknown subjects in motor imagery-based brain-computer interface (MI-BCI). However, the current methods merely achieve balance across different subjects while neglecting inter-class relationships, and the model is easily biased towards distinguishable categories to learn the imbalanced classification boundaries, degrading classification performance on target subjects. Therefore, we propose a subject-class joint gradient alignment-based domain generalization framework (SCGADG). At each iteration, the data of existing subjects is first partitioned into virtual test and training sets simultaneously by subject and category. Both virtual training and test sets consist of samples from some categories of partial subjects and the remaining categories of other subjects. Then, the model is virtually trained and updated using the virtual training set. Subsequently, the loss on the virtual test set is computed to optimize the original model. The SCGADG achieves average accuracies of 67.98% and 84.88% for cross-unknown subjects based on two public MI datasets, respectively. The result shows that by aligning the gradients of the partitioned tasks, the optimized model not only has cross-domain generalization ability but also more precisely captures inter-class relationships, establishing balanced decision boundaries between categories, thereby enhancing classification performance for unseen target subjects.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1742-6588
1742-6596
DOI:10.1088/1742-6596/3079/1/012083