Beyond Hearing: Learning Task-agnostic ExG Representations from Earphones via Physiology-informed Tokenization

Position

Project member in Microsoft Research Asia

Project Goals & Works

1. Built a task-agnostic ExG representation learning pipeline for real-world, free-living sensing settings.
2. Proposed Physiology-informed Multi-band Tokenization (PiMT) to represent ExG signals with physiology-aware token decomposition.
3. Trained models with reconstruction-based objectives to capture robust, transferable features across tasks.
4. Evaluated on DailySense and multiple public ExG benchmarks to validate cross-task generalization.

Key Results

1. Demonstrated consistent performance gains over prior ExG baselines across multiple sensing tasks.
2. Showed strong robustness in unconstrained daily-life data collection settings.
3. Highlighted scalability of earphone-based ExG sensing for practical on-device intelligence.

Links

• Website: https://miil.kaist.ac.kr/projects/beyondhearing
• OpenReview: https://openreview.net/pdf?id=s79tJrxDmt
• arXiv: https://arxiv.org/abs/2510.20853