Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

Bridging Encoding to Decoding: Synthetic fMRI from Brain Foundation Models Improves Neural Decoding

Zhejun Zhang¹, Wenqing Zhou², Han Mengqian², Lin Zhang¹, Lei Li¹; ¹Beijing University of Posts and Telecommunications, ²Beijing University of Post and Telecommunications

Presenter: Zhejun Zhang

Encoding models predict brain activity from stimuli; decoding models invert this mapping—yet the knowledge accumulated in large-scale encoding models has not been leveraged for decoder training. Here we bridge them by using a brain encoding foundation model (TRIBE v2) to synthesize training data for decoder learning. We generate ~5,000 synthetic fMRI responses to natural images, align their distributions with real recordings through noise injection, and pretrain a visual decoder on this augmented set before finetuning on limited real data. This approach yields consistent improvements in retrieval-based decoding across all data regimes (28–42% relative gain, *p* < 0.05), and a shuffled-pairing control confirms that the gains arise from the encoding model's learned stimulus–brain mappings. These results demonstrate that knowledge accumulated in encoding foundation models can be transferred to decoding via synthetic data generation, offering a practical pathway to alleviate data constraints in neural decoding.

Topic Area: Methods, Tools, Theory & Neural Coding