Beyond Trial Averaging: Anchoring Neural and Visual Representations for Few-Repetition Brain-to-Image Retrieval
NEAR uses high-repetition brain-signal centers as anchors to improve image retrieval when only a few neural trials are available.
The paper says current brain-to-image retrieval often depends on averaging many repeated trials per image, sometimes up to 80. Its authors argue that low-repetition failures are not just query noise: the neural query and image representation both align with the high-repetition center, but not directly with each other. Their NEAR framework pulls both the noisy brain signal and each image candidate toward that anchor. Across EEG, MEG, and fMRI datasets, it improved few-repetition retrieval, including 200-way Top-1 gains of 5.7 and 9.3 points on THINGS-EEG2 for one and four repetitions. ArXiv · AI/CL/LG's note
The paper says current brain-to-image retrieval often depends on averaging many repeated trials per image, sometimes up to 80. Its authors argue that low-repetition failures are not just query noise: the neural query and image representation both align with the high-repetition center, but not directly with each other. Their NEAR framework pulls both the noisy brain signal and each image candidate toward that anchor. Across EEG, MEG, and fMRI datasets, it improved few-repetition retrieval, including 200-way Top-1 gains of 5.7 and 9.3 points on THINGS-EEG2 for one and four repetitions. ArXiv · AI/CL/LG's note
score 4