Tokenizing Single-Channel EEG with Time-Frequency Motif Learning

ICLR 2026 Conference Submission5323 Authors

Keywords: EEG, tokenization, representation learning

TL;DR: We propose TFM-Tokenizer, a EEG tokenization framework that encodes single-channel EEG into discrete tokens, improving downstream performance, enhancing existing FMs as a plug-in component, and scaling to other brain signal modalities.

Abstract:

Foundation models are reshaping EEG analysis, yet an important problem of EEG tokenization remains a challenge.
This paper presents TFM-Tokenizer, a novel tokenization framework that learns a vocabulary of time-frequency motifs from *single-channel* EEG signals and encodes them into discrete tokens.
We propose a dual-path architecture with time–frequency masking to capture robust motif representations, and it is model-agnostic, supporting both lightweight transformers and existing foundation models for downstream tasks.
Our study demonstrates three key benefits:
*Accuracy:* Experiments on four diverse EEG benchmarks demonstrate consistent performance gains across both single- and multi-dataset pretraining settings, achieving up to 17% improvement in Cohen’s Kappa over strong baselines.
*Generalization:* Moreover, as a plug-and-play component, it consistently boosts the performance of diverse foundation models, including BIOT and LaBraM.
*Scalability:* By operating at the single-channel level rather than relying on the strict 10–20 EEG system, our method has the potential to be device-agnostic.
Experiments on ear-EEG sleep staging, which differs from the pretraining data in signal format, channel configuration, recording device, and task, show that our tokenizer outperforms baselines by 14%.
A comprehensive token analysis reveals strong class-discriminative, frequency-aware, and consistent structure, enabling improved representation quality and interpretability.
Code is available at https://anonymous.4open.science/r/TFM-Token-FE33.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 5323