ICLR 2026

Beyond Hearing: Learning Task-Agnostic ExG Representations from Earphones via Physiology-Informed Tokenization

Hyungjun Yoon, Seungjoo Lee, Yu Yvonne Wu, Xiaomeng Chen, Taiting Lu, Freddy Yifei Liu, Taeckyung Lee, Hyeongheon Cha, Haochen Zhao, Gaoteng Zhao, Dongyao Chen, Cecilia Mascolo, Sung-Ju Lee, and Lili Qiu

Paper arXiv

NeuroBuds earphone-based ExG sensing: free-living ear-ExG feeds a task-agnostic representation model evaluated across five human senses

Abstract

Electrophysiological (ExG) signals offer valuable insights into human physiology, yet building foundation models that generalize across everyday tasks remains challenging: most ExG recordings are collected in controlled labs with bulky, expensive devices, and task-specific model designs limit generalization across tasks.

We introduce an approach for scalable, task-agnostic ExG monitoring in the wild. With NeuroBuds, an earphone-based ExG prototype capturing near-ear EEG, facial EMG, and EOG, we collected 50 hours of unobtrusive free-living ExG data. At the core of our approach is Physiology-informed Multi-band Tokenization (PiMT), which decomposes ExG signals into 12 physiology-informed tokens and learns robust representations through reconstruction. Experiments on our new DailySense dataset — the first to enable ExG-based analysis across the five human senses — together with four public ExG benchmarks show that PiMT consistently outperforms state-of-the-art methods across diverse tasks.

Key Contributions

NeuroBuds: a lightweight, low-cost earphone prototype that captures near-ear EEG, facial EMG, and EOG for long-term, unobtrusive ExG monitoring.
DailySense: 50 hours of free-living ExG recordings from 22 participants plus 20 hours of task-specific data — the first benchmark spanning the five human senses.
PiMT: physiology-informed multi-band tokenization that keeps representations task-agnostic while exposing task-relevant frequency structure, achieving state-of-the-art results (average F1 of 87.6%) across six DailySense tasks and four public benchmarks.

BibTeX

@inproceedings{yoon2026beyond,
  title     = {Beyond Hearing: Learning Task-Agnostic ExG Representations from Earphones via Physiology-Informed Tokenization},
  author    = {Yoon, Hyungjun and Lee, Seungjoo and Wu, Yu Yvonne and Chen, Xiaomeng and Lu, Taiting and Liu, Freddy Yifei and Lee, Taeckyung and Cha, Hyeongheon and Zhao, Haochen and Zhao, Gaoteng and Chen, Dongyao and Mascolo, Cecilia and Lee, Sung-Ju and Qiu, Lili},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026}
}