Decoupling representation learning for imbalanced electroencephalography classification in rapid serial visual...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Class imbalance significantly hinders electroencephalography (EEG) classification performance in rapid serial visual presentation (RSVP) tasks.
• Current re-balancing strategies distort feature distributions, damaging deep network representational capacity.

Purpose of the Study:

• To propose a novel Decoupling Representation Learning (DRL) model to accurately classify imbalanced RSVP EEG data.
• To separate representation learning and classification to capture discriminative features without distorting original data distributions.

Main Methods:

• A dual-branch architecture for representation learning, minimizing contrastive loss to regularize the feature space.
• A multi-granular information extractor to capture complementary spatial-temporal information from RSVP EEG data.
• Training the classifier with re-balanced data while freezing representation learning parameters.

Main Results:

• The proposed DRL model achieved state-of-the-art performance on two public and one self-conducted RSVP EEG datasets.
• Demonstrated superior accuracy in EEG classification for the RSVP task compared to existing methods.

Conclusions:

• The DRL model offers a generic and effective solution for the class imbalance problem in RSVP EEG classification.
• The study highlights the benefit of multi-granular data exploration for extracting richer spatial-temporal features.