VoxelFormer: Parameter-Efficient Multi-Subject Visual Decoding from fMRI

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Le, Chenqian, Zhao, Yilin, Emami, Nikasadat, Yadav, Kushagra, Liu, Xujin "Chris", Chen, Xupeng, Wang, Yao
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866911149220429824
author Le, Chenqian
Zhao, Yilin
Emami, Nikasadat
Yadav, Kushagra
Liu, Xujin "Chris"
Chen, Xupeng
Wang, Yao
author_facet Le, Chenqian
Zhao, Yilin
Emami, Nikasadat
Yadav, Kushagra
Liu, Xujin "Chris"
Chen, Xupeng
Wang, Yao
contents Recent advances in fMRI-based visual decoding have enabled compelling reconstructions of perceived images. However, most approaches rely on subject-specific training, limiting scalability and practical deployment. We introduce \textbf{VoxelFormer}, a lightweight transformer architecture that enables multi-subject training for visual decoding from fMRI. VoxelFormer integrates a Token Merging Transformer (ToMer) for efficient voxel compression and a query-driven Q-Former that produces fixed-size neural representations aligned with the CLIP image embedding space. Evaluated on the 7T Natural Scenes Dataset, VoxelFormer achieves competitive retrieval performance on subjects included during training with significantly fewer parameters than existing methods. These results highlight token merging and query-based transformers as promising strategies for parameter-efficient neural decoding.
format Preprint
id arxiv_https___arxiv_org_abs_2509_09015
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle VoxelFormer: Parameter-Efficient Multi-Subject Visual Decoding from fMRI
Le, Chenqian
Zhao, Yilin
Emami, Nikasadat
Yadav, Kushagra
Liu, Xujin "Chris"
Chen, Xupeng
Wang, Yao
Computer Vision and Pattern Recognition
Recent advances in fMRI-based visual decoding have enabled compelling reconstructions of perceived images. However, most approaches rely on subject-specific training, limiting scalability and practical deployment. We introduce \textbf{VoxelFormer}, a lightweight transformer architecture that enables multi-subject training for visual decoding from fMRI. VoxelFormer integrates a Token Merging Transformer (ToMer) for efficient voxel compression and a query-driven Q-Former that produces fixed-size neural representations aligned with the CLIP image embedding space. Evaluated on the 7T Natural Scenes Dataset, VoxelFormer achieves competitive retrieval performance on subjects included during training with significantly fewer parameters than existing methods. These results highlight token merging and query-based transformers as promising strategies for parameter-efficient neural decoding.
title VoxelFormer: Parameter-Efficient Multi-Subject Visual Decoding from fMRI
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2509.09015