← Paper Digest

GALA: Gallery-Level Alignment for Real-Life Vision Decoding from EEG

Philippe Forest, Giulia Lioi, Mai Pham Quyen, Alix Lamouroux, Javen Shi, Adrien Merlini

November 2026

TL;DR Calibrate an EEG-to-image decoder to a new user without any labels, by registering the user's growing pool of unlabeled EEG embeddings onto the known gallery of candidate images. On THINGS-EEG-2 with 200 images, top-1 accuracy for an unseen subject rises from 28% to 67% over 53 minutes of use.

What they do

EEG decoders that match a brain response to one of a set of images reach over 80% top-1 accuracy within a subject but about 32% on an unseen subject, and adapting them usually needs labeled calibration data. A deployed BCI, however, knows its gallery of possible outcomes and collects unlabeled EEG as soon as it is used. The authors define incremental transductive calibration: after every query, the decoder is recalibrated on all unlabeled EEG received so far plus the gallery, and each query is decoded before it joins that buffer. GALA is a label-free calibration method for this setting.

How it works

  • A frozen EEG encoder (ATM, trained on 9 source subjects) maps EEG into the space of InternViT-6B image features.
  • Calibration registers the cloud of EEG embeddings onto the cloud of gallery embeddings: shrinkage whitening, CSLS scores against hub images, then alternating Sinkhorn correspondences and an orthogonal Procrustes rotation.
  • How often each gallery image is queried is estimated from the buffer, so unequal and partial query streams are handled.
  • The zero-shot decoder acts as a prior, so with few queries the correction stays near identity. One update takes 15 ms on one CPU core with 400 queries.

Results

  • THINGS-EEG-2, leave-one-subject-out, 200 images, 20 repetitions per query, uniform stream: GALA reaches 37.2% after 1 min, 59.1% at 20 min and 67.3% at 53 min. Zero-shot stays near 28%; the next best, SCORE with the ATM encoder, ends at 53.4%.
  • Across 60 settings of gallery size, repetitions and stream type, GALA is the most accurate label-free method over the last 30% of every stream, by 6.6 points on average.
  • Standard batch test: top-1 rises from 32.2% to 79.1%; a plain one-to-one Hungarian assignment already gives 63.3%.
  • Without the prior, accuracy at 20 min collapses to 4.0%; without the rotation, 39.1%.

Limitations

Gains depend on the encoder: with LaBraM, SCORE is up to 1.7 points better between 5 and 20 minutes, and on Alljoined-1.6M and THINGS-MEG every method stays within 4 points of zero-shot. The evaluation replays offline recordings in which each query averages many repetitions of one image, unlike a closed-loop BCI with single-trial queries.

Figures

From Figure 1: Evaluation settings. In zero-shot evaluation each query is decoded alone; in batch evaluation the 200 ...
From Figure 1: Evaluation settings. In zero-shot evaluation each query is decoded alone; in batch evaluation the 200 test queries are calibrated together and each image is queried once. In incremental transductive calibration, queries arrive one at a time with unequal frequencies, some images are never queried, and each calibration uses all queries received so far.
From Figure 2: GALA. The unlabeled EEG embeddings of an unseen subject are whitened, compared with the gallery throug...
From Figure 2: GALA. The unlabeled EEG embeddings of an unseen subject are whitened, compared with the gallery through cross-domain similarity local scaling, and registered onto it by a Sinkhorn normalization with estimated image frequencies and a soft Procrustes alignment. Each new query triggers one pass of this loop.
From Figure 3: Top-1 accuracy of incoming queries on THINGS-EEG-2 (10 subjects, 20 repetitions per query) against sti...
From Figure 3: Top-1 accuracy of incoming queries on THINGS-EEG-2 (10 subjects, 20 repetitions per query) against stimulus time, for the uniform, skewed and partial streams. Each query is decoded before it is used for calibration. GALA is compared with SCORE (ATM), SATTC and the zero-shot decoder.
Abstract

Neural decoding of visual experience from electroencephalography (EEG) has achieved strong performance in intra-subject settings, but current methods fail to generalize across individuals: on the standard 200-way benchmark THINGS-EEG-2, top-1 accuracy drops from above 80% within subjects to about 32% for an unseen subject. Adapting a pre-trained encoder to new subjects and sessions typically requires labeled data, a bottleneck whose collection time severely limits practical device deployment. In a practical Brain-Computer Interface (BCI), however, the possible outcomes are known in advance and form a gallery of candidates, and the unlabeled EEG of the user accumulates as soon as the device is used. We propose a new setting, incremental transductive calibration, in which the decoder is calibrated regularly on all the unlabeled EEG received so far and on the known gallery, without any labels. This setting is closer to the conditions of a deployed BCI than the zero-shot and batch evaluations in current use. We evaluate it with a protocol in which queries arrive one at a time, with unknown and possibly unequal frequencies, and each query is decoded before it is used for calibration. In this setting, we introduce GALA : GAllery-Level Alignment, a label-free calibration method which registers the unlabeled EEG embeddings onto the gallery through shrinkage whitening, hubness correction and a Sinkhorn--Procrustes alignment with query frequencies estimated from the buffer, updated after every query. On THINGS-EEG-2, with a gallery of 200 images, it raises the top-1 accuracy on a new subject from 28% to 67% over 53 minutes of use. Across 60 evaluation settings that vary the gallery size, the query frequencies and the number of repetitions averaged per query, GALA is the most accurate label-free method in every setting, 6.6 points above the best competing method on average at the end of the stream.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.