Multi-Scale Masked Autoencoders for Cross-Session Emotion Recognition

Area of Science:

• Neuroscience and Artificial Intelligence
• Brain-Computer Interfaces (BCIs)
• Machine Learning for Affective Computing

Background:

• Affective brain-computer interfaces (aBCIs) leverage electroencephalogram (EEG) for emotion recognition.
• Challenges include time-consuming data annotation, individual differences, non-stationarity, and noise in EEG.
• Developing subject-specific, cross-session emotion recognition models remains difficult.

Purpose of the Study:

• To propose a unified pre-training framework, Multi-Scale Masked Autoencoders (MSMAE), to address challenges in EEG-based emotion recognition.
• To extract noise-robust, subject-invariant, and temporal-invariant features from large-scale unlabeled EEG data.
• To enable subject-specific cross-session emotion recognition through fine-tuning with minimal labeled data.

Main Methods:

• Utilized a multi-scale masked autoencoder (MSMAE) framework for pre-training on large-scale unlabeled EEG signals.
• Implemented multi-scale representation to capture diverse EEG signal aspects.
• Employed an improved masking mechanism for channel-level representation and invariance learning for spatial-level representation to minimize variances.

Main Results:

• The MSMAE framework successfully extracted noise-robust, subject-invariant, and temporal-invariant features.
• Demonstrated remarkable ability in decoding emotional states from different EEG sessions.
• Achieved stable and superior performance compared to baseline methods on SEED and SEED-IV datasets.

Conclusions:

• The proposed MSMAE framework effectively overcomes key challenges in subject-specific cross-session emotion recognition.
• MSMAE provides a robust and efficient method for feature extraction from EEG data.
• This approach significantly advances the development of practical affective brain-computer interfaces.