EEGDancer predicts emotions from brain signals with RL-optimized latent spaces
New AI framework maps EEG to emotions using reinforcement learning, beating existing methods on 3 datasets.
EEGDancer tackles the challenge of continuous emotion prediction from EEG signals by moving beyond point-wise regression. The framework first uses a causal spatiotemporal VQ-VAE to learn a discrete-continuous emotional latent space, converting noisy high-dimensional EEG data into structured emotional prototypes. A Transformer-based masked modeling strategy then captures long-range temporal dependencies and emotional dynamics over time.
Finally, continuous prediction is reframed as a sequential decision-making problem and optimized using Soft Actor-Critic (SAC) reinforcement learning, which smooths emotional trajectories at the sequence level rather than frame-by-frame. Tested on the SEED, SEED-IV, and Long-Term Naturalistic Emotion datasets, EEGDancer consistently outperforms existing machine learning and deep learning baselines. The 51-page paper includes 9 figures and 13 tables, with ablation studies validating each component's contribution.
- EEGDancer uses a causal spatiotemporal VQ-VAE to learn structured emotional prototypes from raw EEG signals.
- A Transformer-based masked modeling strategy captures long-range temporal dependencies in emotional states.
- Soft Actor-Critic (SAC) reinforcement learning optimizes prediction trajectories at the sequence level, improving accuracy over frame-wise methods.
Why It Matters
Enables more accurate, real-time emotion tracking from brain signals for mental health and human-computer interaction applications.