[Paper Review] Ultra Efficient Transfer Learning with Meta Update for Cross Subject EEG Classification

Tiehang Duan, Mihir Chauhan+3 more|arXiv (Cornell University)|Mar 13, 2020

EEG and Brain-Computer Interfaces24 references17 citations

TL;DR

This paper proposes MUPS-EEG, a meta-learning-based transfer learning method for cross-subject EEG classification that enables fast adaptation to new subjects with minimal labeled data while preserving knowledge of previously learned subjects. By leveraging meta-learned representations and a meta-update strategy, MUPS-EEG achieves state-of-the-art performance on two public EEG datasets, outperforming prior methods by up to 3.4% in accuracy and 1.5% in AUC-ROC.

ABSTRACT

The pattern of Electroencephalogram (EEG) signal differs significantly across different subjects, and poses challenge for EEG classifiers in terms of 1) effectively adapting a learned classifier onto a new subject, 2) retaining knowledge of known subjects after the adaptation. We propose an efficient transfer learning method, named Meta UPdate Strategy (MUPS-EEG), for continuous EEG classification across different subjects. The model learns effective representations with meta update which accelerates adaptation on new subject and mitigate forgetting of knowledge on previous subjects at the same time. The proposed mechanism originates from meta learning and works to 1) find feature representation that is broadly suitable for different subjects, 2) maximizes sensitivity of loss function for fast adaptation on new subject. The method can be applied to all deep learning oriented models. Extensive experiments on two public datasets demonstrate the effectiveness of the proposed model, outperforming current state of the arts by a large margin in terms of both adapting on new subject and retain knowledge of learned subjects.

Motivation & Objective

• Address the challenge of significant EEG pattern variability across subjects in brain-computer interface (BCI) systems.
• Enable fast and accurate adaptation of EEG classifiers to new, previously unseen subjects with minimal labeled data.
• Prevent catastrophic forgetting of previously learned subject knowledge during adaptation.
• Develop a generalizable, computationally efficient transfer learning framework applicable to any deep learning-based EEG classifier.
• Facilitate real-world deployment of BCI systems with dynamic subject changes by improving cross-subject generalization.

Proposed method

• The method employs a two-stage meta-learning framework: meta representation learning on source subjects and meta adaptation on target subjects.
• Meta representation learning extracts versatile, subject-invariant features from multiple source subjects to improve generalization.
• Meta adaptation uses a few gradient steps on target subject data to rapidly fine-tune the model while preserving knowledge of source subjects.
• The meta-update mechanism optimizes for fast adaptation by maximizing the sensitivity of the loss function to small updates on new data.
• The approach is model-agnostic and can be applied to any deep learning architecture used for EEG classification.
• The optimization is formulated as a hierarchical Bayesian inference problem, approximating the posterior over meta-parameters to balance adaptation and knowledge retention.

Experimental results

Research questions

• RQ1 Can a meta-learning approach significantly reduce the amount of labeled target data required for effective cross-subject EEG classification?
• RQ2 How well can a model retain knowledge of previously learned subjects while adapting to a new subject?
• RQ3 Does meta-learned representation learning improve generalization across diverse EEG signal patterns across subjects?
• RQ4 How does the proposed meta-update strategy compare to classic transfer learning and domain adaptation methods in terms of accuracy and robustness?
• RQ5 What is the minimum amount of target subject data required for convergence and optimal performance?

Key findings

• On the BCI IV-2a dataset, MUPS-EEG achieved at least 2.2% higher accuracy and 1.3% higher AUC-ROC compared to state-of-the-art methods.
• For the DEAP dataset, MUPS-EEG outperformed all baselines by at least 3.4% in accuracy and 1.5% in AUC-ROC.
• MUPS-EEG classified 7 out of 9 subjects on BCI IV-2a to above 70% accuracy, meeting the threshold for practical BCI application.
• The model demonstrated superior knowledge retention, achieving 1.6% higher average accuracy and 0.5% higher average AUC-ROC on source subjects compared to baselines on the BCI IV-2a dataset.
• On the more challenging DEAP dataset, MUPS-EEG achieved a 2.6% gain in average accuracy and 0.7% gain in average AUC-ROC over the next best method.
• Performance converged with only 2 minutes of target data on BCI IV-2a and 5 minutes on DEAP, indicating high data efficiency.