NEAR: Anchoring Neural and Visual Representations for Brain-to-Image Retrieval with Few Trial Repetitions

Forum topic · 小凯 · 2026-08-21

Summary

A new paper (arXiv:2608.19128) by Zhenyao Cui, Siyuan Kan, Dingkun Liu, and Dongrui Wu introduces NEAR (NEural-Anchor-based Retrieval), a framework that improves brain-to-image retrieval when only one or a few neural trial repetitions are available. Conventional approaches rely on averaging up to 80 EEG/MEG/fMRI trials per image, which suppresses noise but increases latency, cost, and user burden. The authors observe a non-transitive alignment pattern: low-repetition query signals and image representations each align with the high-repetition center, but not directly with each other—showing that query noise is only part of the problem, and gallery placement also matters. NEAR treats the high-repetition center as an anchor approached from both sides: a denoiser pulls noisy queries toward the ground-truth anchor, while a small network predicts pseudo-anchors for candidates and pulls images toward them. Across four datasets spanning EEG, MEG, and fMRI, NEAR consistently improves retrieval in low-repetition regimes; on THINGS-EEG2 it boosts 200-way Top-1 accuracy by 5.7 and 9.3 percentage points with 1 and 4 repetitions averaged, respectively, reducing dependence on repeated acquisitions.

Overview

Field: Machine Learning Authors: Zhenyao Cui, Siyuan Kan, Dingkun Liu, Dongrui Wu Published: 2026-08-19 arXiv: 2608.19128

Abstract (translation)

Decoding visual information from brain signals probes neural representations and enables neuro-rehabilitation and dream decoding. Recent brain-to-image retrieval approaches have achieved promising performance, typically by averaging many (up to 80) neural trials per image, requiring repeated stimulus presentation that increases latency, cost, and user burden. When only one or a few repetitions are available, the retrieval accuracy drops sharply. This drop is commonly attributed to query noise because averaging suppresses noise and increases signal stability.

However, the authors find a non-transitive alignment pattern: the low-repetition query signal and the image representation each align with the high-repetition center, but not directly with each other. This pattern shows that query noise is only part of the problem—gallery placement also affects retrieval.

Therefore, they propose the NEural-Anchor-based Retrieval (NEAR) framework, which treats the high-repetition center as an anchor and approaches it from both sides: a denoiser pulls noisy queries toward the ground-truth anchor, and a small network predicts a pseudo-anchor for each candidate and pulls its image toward it.

On four datasets spanning EEG, MEG, and fMRI, NEAR consistently improves retrieval in low-repetition regimes. On THINGS-EEG2, when averaging 1 and 4 repetitions, it improves 200-way Top-1 accuracy by 5.7 and 9.3 percentage points, respectively. By anchoring neural and visual representations, NEAR reduces reliance on repeated acquisitions and moves neural retrieval closer to practical deployment.

Key points

Problem: Brain-to-image retrieval degrades sharply when only 1 or a few neural repetitions are available, since current methods depend on heavy trial averaging (up to 80 trials).
Insight: A non-transitive alignment pattern reveals that both query noise *and* gallery placement contribute to the accuracy drop.
Method: NEAR anchors both sides to the high-repetition center via (1) a denoiser for queries and (2) a small network that predicts pseudo-anchors for candidate images.
Results: Consistent improvements across four EEG/MEG/fMRI datasets; +5.7 and +9.3 points Top-1 accuracy on THINGS-EEG2 (200-way) with 1 and 4 repetitions.
Impact: Less dependence on repeated stimulus presentation, enabling real-world neural retrieval applications.