MIN2Net: End-to-End Multi-Task Learning for Subject-Independent Motor Imagery EEG Classification

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery (MI)-based brain-computer interfaces (BCIs) leverage electroencephalography (EEG) for non-invasive control.
• Subject-specific EEG rhythms and temporal variations pose challenges for robust BCI performance, especially in subject-independent scenarios.

Purpose of the Study:

• To introduce MIN2Net, a novel end-to-end multi-task learning framework for subject-independent MI-based BCI.
• To improve EEG classification performance by learning compact and discriminative latent representations.

Main Methods:

• MIN2Net integrates deep metric learning with a multi-task autoencoder.
• The model learns latent representations from EEG data and performs classification concurrently.
• This approach aims to reduce preprocessing complexity.

Main Results:

• MIN2Net achieved significant performance improvements in EEG classification.
• Experimental results demonstrated superior performance over state-of-the-art techniques in a subject-independent manner.
• An F1-score improvement of 6.72% on the SMR-BCI dataset and 2.23% on the OpenBMI dataset was observed.

Conclusions:

• MIN2Net effectively enhances discriminative information within the latent representation of EEG data.
• The study highlights the potential of MIN2Net for developing practical MI-BCI applications for new users without requiring prior calibration.