MILFNet: Multiple Instance Learning-Based FastSlow Network with Multitask Autoencoder for Cross-Subject EEG Emotion Recognition

As humanoid robots and artificial intelligence technologies advance, accurately recognizing human emotions has become increasingly crucial in human-computer interactions. This capability enhances user experience across various applications such as virtual reality, education, and healthcare. Electroencephalography (EEG) signals offer a noninvasive and objective means of measuring affective brain states, providing direct insights into neural activities associated with emotions. In this context, EEG-based emotion recognition presents a significant measurement challenge of extracting reliable affective markers from complex neural signals with high intersubject variability. Addressing these issues, we propose a multiple instance learning-based FastSlow network (MILFNet). MILFNet adopts a weakly supervised learning framework with signal-level labels to mitigate label noise and incorporates a two-phase multitask autoencoder to reduce intersubject variability. We evaluated the model under subject-independent environments using three publicly available datasets: SEED, SEED-IV, and Dreamer. MILFNet significantly outperforms the state-of-the-art algorithms, achieving accuracies of 97.25% on SEED, 87.68% on SEED-IV, and 84.52% for valence and 84.97% for arousal on Dreamer. These results demonstrate MILFNet's robustness to label noise and intersubject differences, as well as its potential as a reliable framework for detecting affective brain responses. This positions MILFNet as a promising tool for advancing emotion-aware instrumentation systems and practical affective computing applications.