Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
SynBrain: Enhancing Visual-to-fMRI Synthesis via Probabilistic Representation Learning
Authors: Weijian Mai, Jiamin Wu, Yu Zhu, Zhouheng Yao, Dongzhan Zhou, Andrew Luo, Qihao Zheng, Wanli Ouyang, Chunfeng Song
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that Syn Brain surpasses stateof-the-art methods in subject-specific visual-to-f MRI encoding performance. Furthermore, Syn Brain adapts efficiently to new subjects with few-shot data and synthesizes high-quality f MRI signals that are effective in improving data-limited f MRI-to-image decoding performance. Beyond that, Syn Brain reveals functional consistency across trials and subjects, with synthesized signals capturing interpretable patterns shaped by biological neural variability. |
| Researcher Affiliation | Academia | 1Shanghai Artificial Intelligence Laboratory 2The University of Hong Kong 3The Chinese University of Hong Kong 4Shanghai Innovation Institute |
| Pseudocode | Yes | Algorithm 1 Brain VAE Architecture 1: Input: f MRI signal x, CLIP visual representation z CLIP 2: Encoder: 3: h = Encoder(x) Hierarchical Conv+Attn backbone 4: Pre-projector (MLP) for Mean or Log Var: 5: h = Layer Norm(h) GELU 6: h = Linear(4096 d) Layer Norm GELU d = 2048 7: h = Linear(d d) Layer Norm GELU 8: or log 2 = Linear(d 1664) 9: Sampling: 10: z N(, 2) 11: Post-Projector (MLP): 12: z = Layer Norm(z) GELU 13: z = Linear(1664 d) Layer Norm GELU d = 2048 14: z = Linear(d d) Layer Norm GELU 15: h = Linear(d 4096) 16: Decoder: 17: x = Decoder(h , Vs) Vs denotes the voxel count for subject S 18: Loss: 19: L = MSELoss(x, x) + KLWeight KL(, 2) + CLIPWeight CLIPLoss(z, z CLIP) 20: return L |
| Open Source Code | Yes | Our code is available at https://github.com/Michael Maiii/Syn Brain. |
| Open Datasets | Yes | We conduct experiments on the Natural Scenes Dataset (NSD) [2], a large-scale f MRI dataset in which 8 subjects viewed natural images from the COCO dataset [29] across approximately 40 hours of scanning. |
| Dataset Splits | Yes | For each subject, we use 9,000 unique images for training and evaluate on a shared set of 1,000 test images, each presented across 3 trials to account for response variability. |
| Hardware Specification | Yes | Our model is implemented and trained on 4 NVIDIA A100 GPUs (40GB memory per GPU), with training completed within 2 hours. |
| Software Dependencies | No | We adopt the pretrained Open CLIP Vi T-big G/14 model [43] as a frozen visual encoder to extract semantic embeddings from visual stimuli. Our model is implemented and trained on 4 NVIDIA A100 GPUs (40GB memory per GPU), with training completed within 2 hours. |
| Experiment Setup | Yes | We train Brain VAE using the Adam W optimizer with (1, 2) = (0.9, 0.999), a learning rate of 1 104, weight decay of 0.05, and a batch size of 24. We apply early stopping to prevent overfitting. The S2N Mapper is optimized with identical hyperparameters and trained for 50,000 steps. |