BCIwiki (bciwiki.com) — A lightweight adapter that aligns EEG and MEG signals with every patch token from an intermediate layer of a vision encoder reached 95.4% within-subject Top-1 accuracy on the THINGS-EEG2 dataset, Yuhui Jin, Yonghao Song and Bingchuan Liu reported in a preprint posted to arXiv on September 19, 2026. The work has not been peer reviewed.
Most existing M/EEG visual decoding methods align brain signals with a single global embedding from a pretrained visual encoder, the authors wrote, leaving unused the intermediate patch representations that preserve finer visual information. Their Residual Patch Adapter (RPA) feeds all patch tokens from an intermediate layer of a Vision Transformer (ViT) into the alignment. Ablation tests showed that pooling or masking the patch tokens degraded the learned representation, while the CLS token added little unique information.
A series of 6 quantitative feature analyses indicated that both high-level semantics and low-level features such as color and texture are essential for EEG-to-image alignment. Under current protocols, the system achieved Top-1 accuracies of 95.4% within-subject and 35.5% cross-subject on THINGS-EEG2, and 65.2% and 6.7% on THINGS-MEG. The authors describe the results as state-of-the-art on both datasets.
The approach also worked with alternative brain encoders, including pretrained EEG foundation models, according to the paper, and RPA offers a plug-and-play interface that can swap in convolution, attention or ConvNeXt modules. The authors say the findings set design principles for brain-image alignment in non-invasive brain-computer interfaces (BCIs).