[Paper Review] Transfer Learning in Brain-Computer Interfaces

Vinay Jayaram, Morteza Alamgir+3 more|arXiv (Cornell University)|Dec 1, 2015

EEG and Brain-Computer Interfaces41 references118 citations

TL;DR

This paper proposes a novel transfer learning framework for brain-computer interfaces (BCIs) that enables knowledge transfer across subjects and sessions by learning shared decision boundary priors in any arbitrary feature space, using a multitask regression approach with fused sparsity. It outperforms standard data pooling and prior transfer methods on both motor-imagery and ALS patient data, achieving high accuracy with minimal training trials by exploiting shared structure in EEG signals.

ABSTRACT

The performance of brain-computer interfaces (BCIs) improves with the amount of available training data, the statistical distribution of this data, however, varies across subjects as well as across sessions within individual subjects, limiting the transferability of training data or trained models between them. In this article, we review current transfer learning techniques in BCIs that exploit shared structure between training data of multiple subjects and/or sessions to increase performance. We then present a framework for transfer learning in the context of BCIs that can be applied to any arbitrary feature space, as well as a novel regression estimation method that is specifically designed for the structure of a system based on the electroencephalogram (EEG). We demonstrate the utility of our framework and method on subject-to-subject transfer in a motor-imagery paradigm as well as on session-to-session transfer in one patient diagnosed with amyotrophic lateral sclerosis (ALS), showing that it is able to outperform other comparable methods on an identical dataset.

Motivation & Objective

• To address the challenge of performance degradation in BCIs due to variability in EEG signal distributions across subjects and sessions.
• To develop a transfer learning method that generalizes across subjects and sessions without relying solely on spatial filtering or domain adaptation.
• To enable efficient, low-latency BCI calibration by leveraging prior data from multiple subjects or sessions.
• To provide a flexible framework compatible with any feature space and objective function, enhancing adaptability to new BCI paradigms.

Proposed method

• Proposes a multitask learning framework that jointly learns decision boundaries across multiple subjects or sessions using a shared prior distribution over weight vectors.
• Introduces a fused sparsity-inducing regularization (FD regression) that encourages shared spatial and frequency weight patterns across tasks.
• Employs a hierarchical Bayesian model to infer priors on weight vectors, allowing transfer from prior subjects to new ones.
• Uses an iterative optimization scheme to update priors and task-specific weights, with convergence accelerated by the FD regularization.
• Applies the framework to both motor-imagery and a novel cognitive BCI paradigm, using EEG features without prior constraints on electrode selection.
• Supports both subject-to-subject and session-to-session transfer by modeling shared structure in the decision boundary space.

Experimental results

Research questions

• RQ1 Can transfer learning improve BCI performance when training data is limited by reducing calibration time across subjects and sessions?
• RQ2 How does a shared prior over decision boundary weights compare to simple data pooling or domain adaptation in handling inter-subject and inter-session EEG variability?
• RQ3 Can a multitask learning framework with fused sparsity generalize effectively across diverse EEG feature spaces and BCI paradigms?
• RQ4 Does the proposed FD regression method achieve faster convergence and better performance than standard multitask learning on high-dimensional EEG data?
• RQ5 To what extent can this framework be used to rapidly evaluate new BCI paradigms before full subject-specific calibration?

Key findings

• The proposed framework outperformed standard data pooling and prior transfer methods on subject-to-subject transfer in a motor-imagery BCI, achieving comparable or higher accuracy with significantly fewer training trials.
• On session-to-session transfer in an ALS patient, the method achieved high classification accuracy with minimal retraining, demonstrating robustness to long-term EEG signal variability.
• The FD regression variant converged orders of magnitude faster than non-FD counterparts due to a more favorable ratio of trials to features, despite an additional optimization loop.
• The method demonstrated strong performance on high-dimensional EEG feature spaces, showing robustness even when many features contributed to classification.
• The framework enabled effective knowledge transfer without requiring prior assumptions about optimal electrode locations or signal bands, making it adaptable to new paradigms.
• The approach was shown to be extensible to other objective functions, suggesting broad applicability beyond the tested regression setup.