Research map / Architectures and efficiency / Attention and transformers
Can Natural Image Autoencoders Compactly Tokenize fMRI Volumes for Long-Range Dynamics Modeling?
Peter Yongho Kim, Juhyeon Park, Jungwoo Park, Jubin Choi, Jungwoo Seo, Jiook Cha, Taesup Moon
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace SwiFT with TABLeT (27 tokens per volume, T=256) | voxel-based fMRI models · ADHD-200 Diagnosis | accuracy: 63.9 -> 65.8 | Improved |
| TABLeT versus SwiFT on UKB sex accuracy | TABLeT · UKB Sex | accuracy: 98.1 -> 97.7 | No clear effect |
| TABLeT versus SwiFT on HCP age | TABLeT · HCP Age | MAE: 0.699 -> 0.705 | No clear effect |
| add masked token modeling pre-training on UKB before fine-tuning (ablation) | TABLeT · HCP Age | Pearson rho: 0.473 -> 0.552 | Improved |
| add masked token modeling pre-training (ablation) | TABLeT · HCP Sex | AUC: 0.987 -> 0.986 | No clear effect |
| use off-the-shelf 2D natural-image DCAE instead of 3D DCAE trained on fMRI (ablation) | TABLeT tokenizer · HCP Sex | accuracy: 92.2 -> 93.8 | Improved |
| use 2D DCAE instead of 3D DCAE (ablation) | TABLeT tokenizer · ADHD-200 Diagnosis | F1: 0.644 -> 0.630 | Mixed |
| fine-tune the 2D DCAE on fMRI (ablation) | 2D DCAE · fMRI reconstruction | reconstruction quality | Hurt |
| aggregate three slicing axes instead of a single axis (ablation) | TABLeT tokens · UKB Age | MSE: 0.369 -> 0.340 (best single axis sagittal) | Improved |
| aggregate three axes instead of coronal only (ablation) | TABLeT tokens · HCP Age | MSE: 0.855 -> 0.773 | No clear effect |
| increase number of input frames T (50 to 256) (ablation) | TABLeT · HCP Intelligence and ADHD-200 | MAE / AUC | Improved |
| use TABLeT instead of SwiFT at T=50 | voxel-based model · Peak memory (batch size 4) | memory: 7.33x more memory efficient | Improved |
Closest papers by meaning
- FAST-Brain: A Flow-Aligned Spatio-Temporal Surrogate Brain Model (NeurIPS 2026 · similarity 0.86)
- Brain Harmony: A Multimodal Foundation Model Unifying Morphology and Function into 1D Tokens (NeurIPS 2025 · similarity 0.86)
- A Compact Implicit Neural Representation for Efficient Storage of Massive 4D Functional Magnetic Resonance Imaging (AAAI 2025 · similarity 0.85)
- A scalable generative model for dynamical system reconstruction from neuroimaging data (NeurIPS 2024 · similarity 0.85)
- BrainWorld: A Structural-Prior-Conditioned Generative Model for Whole-Brain 4D fMRI Dynamics (NeurIPS 2026 · similarity 0.85)
- BaRISTA: Brain Scale Informed Spatiotemporal Representation of Human Intracranial Neural Activity (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
