Transfer Learning for P300 Brain-Computer Interfaces by Joint Alignment of Feature Vectors

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interface (BCI) systems often require subject-specific training, limiting their generalizability.
• Transfer learning aims to improve BCI performance by leveraging data from multiple subjects or sessions.
• Existing methods struggle with efficiently aligning diverse BCI datasets.

Purpose of the Study:

• To introduce a novel many-to-many transfer learning method called group learning.
• To develop an extension, fast alignment, for many-to-one transfer learning to previously aligned domains.
• To evaluate the proposed group alignment algorithm (GALIA) on BCI data for classification performance and computational cost.

Main Methods:

• Group learning algorithm (GALIA) utilizes cyclic approximate joint diagonalization (AJD) for joint domain alignment.
• GALIA finds linear transformations to align feature vectors across multiple domains.
• Fast alignment enables transfer learning to new, unseen domains without retraining.

Main Results:

• Group learning and fast alignment significantly improved classification accuracy compared to subject-specific models (average improvement: 2.12±1.88%).
• Performance gains were observed across six public P300 BCI databases (333 sessions, 177 subjects).
• The methods demonstrated effective many-to-many and many-to-one transfer learning for non-clinical BCI data.

Conclusions:

• Group learning provides effective many-to-many transfer learning for BCI, creating a single, robust model.
• Fast alignment extends group learning for efficient many-to-one transfer to unseen domains.
• The proposed methods offer improved classification accuracy and reduced training requirements for BCI systems.