SATTC

Qunjie Huang, Weina Zhu
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026

Structure-Aware Label-Free Test-Time Calibration for Cross-Subject EEG-to-Image Retrieval

Overview

SATTC is a case study within a broader research agenda on reliable inference from generalizable neural representations. Cross-subject EEG-to-image retrieval provides a concrete setting in which representations learned from source subjects must support visual-semantic decoding for a previously unseen person.

In this setting, subject-specific distribution shifts can distort the geometry of the retrieval space even when the pretrained encoder captures useful neural representations. SATTC studies what can be recovered at inference time: it uses the structure of unlabeled test-time representations to recalibrate similarity relationships without retraining the encoder.

Research question

How can learned neural representations remain reliable under subject shift when target labels are unavailable and the pretrained encoder is fixed?

SATTC approaches this question through retrieval-space calibration. It complements representation learning by examining the reliability boundary that remains after a neural encoder has already been trained.

Approach

SATTC operates as a plug-and-play calibration stage on top of pretrained EEG-to-image retrieval models.

Its central ideas are:

• Structure-aware calibration: exploits relationships among unlabeled test-time representations to estimate reliable local structure.
• Adaptive similarity correction: recalibrates cross-modal similarities to reduce distortions caused by subject shift and retrieval hubness.
• Consistency-aware refinement: uses reciprocal and neighborhood-level relationships to improve the reliability of retrieved candidates.
• Encoder-agnostic inference: can be applied without retraining or modifying the underlying EEG encoder.

Why it matters

Transferable representation learning is the foundation of cross-subject neural decoding, but inference can still fail when subject shift distorts the geometry used to match brain and image representations. SATTC isolates this downstream failure mode and studies calibration as one complementary mechanism for reliable decoding.

This perspective suggests that reliable neurotechnology may require both:

1. strong pretrained neural representations, and
2. principled inference mechanisms that expose and address residual distribution shift.

Key takeaways

• Cross-subject retrieval errors arise not only from representation shift, but also from distorted retrieval geometry.
• Unlabeled test-time structure can provide useful signals for calibration.
• Test-time calibration can be separated from encoder design and applied as a plug-and-play component.
• Reliable inference complements, but does not substitute for, transferable representation learning.

Resources

Citation

@inproceedings{huang2026sattc,
  title     = {SATTC: Structure-Aware Label-Free Test-Time Calibration for Cross-Subject EEG-to-Image Retrieval},
  author    = {Huang, Qunjie and Zhu, Weina},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year      = {2026}
}