EEG foundation model fine-tuning sheds subject variation: content and style split, cross-subject decoding steadier
Synopsis
EEG foundation model fine-tuning gains a plug-and-play adapter, DEFT. Pretrained embeddings previously retained substantial subject-related variation, adding burden to the downstream classifier in cross-subject settings. DEFT splits pretrained tokens into a task-oriented content branch and a subject-consistent style branch, then predicts from the content branch, using recursive Gaussian refinement, temporal-variance regularization, subject-aware contrastive learning, and reconstruction to instantiate this factorize-before-classify design. Across multiple EEG datasets and four foundation model backbones it improves matched fine-tuning baselines, and ablations support the intended content/style branch organization.
Interpretation
Fine-tuning of EEG foundation models gains a path that splits pretrained representations into a task-content branch and a subject-style branch, with downstream prediction taken only from the content branch. Earlier fine-tuning used pretrained embeddings directly in the downstream classifier, and those embeddings still retained substantial subject-related variation, which became extra burden in cross-subject settings. Across multiple EEG datasets and four EEG foundation model backbones, the adapter improves on matched fine-tuning baselines.
The mechanism is factorize-before-classify: pretrained tokens are separated into task-oriented content and subject-consistent style, and predictions come from the content branch. The design is instantiated by recursive Gaussian refinement, temporal-variance regularization, subject-aware contrastive learning, and reconstruction, and it attaches to existing fine-tuning pipelines as a plug-and-play adapter. Further ablations and analyses support the intended organization of the content and style branches.
Perspective
The adapter targets the fine-tuning stage of EEG foundation models, for downstream classification in cross-subject settings, and is validated on multiple EEG datasets and four foundation model backbones. Readers focused on EEG decoding pipelines can watch how it performs as a plug-and-play component attached to existing fine-tuning workflows.
The material is summary-level and gives no specific metric values or improvement magnitudes per dataset, nor the exact cross-subject split. Ablations and analyses support the content/style branch organization, but whether the branches stay stable across all backbones and tasks is not further detailed.