Train on spheres, predict brains: GNN transfers zero-shot to cortical folding
bravo_abad · x · 2026-09-21
Zhao et al. tackle cortical folding prediction with a "learn on simple systems, transfer to the real one" strategy: realistic folding simulations are expensive and longitudinal MRI data scarce.
They build dense simulation libraries of growing spheres and ellipsoids governed by the same nonlinear elasticity and mechanical instabilities that drive cortical folding, train a graph neural network on these cheap deformations, then transfer the model zero-shot to human brain surfaces reconstructed from MRI. A practical recipe for ML on systems too complex to simulate and too data-starved to train on directly.
More from Research
- AMD publishes educational GEMM ladder for Helios GPUs: 432GB HBM4, 23TB/s bandwidth — simran_s_arora · 2026-09-22
- Deep learning models learn chemistry from less data; fine-tuning on pharma ELN works — marwinsegler · 2026-09-22
- Microsoft team's retroChimera in Nature: ensembling diverse models for chemist-aligned retrosynthesis — marwinsegler · 2026-09-22
- retroChimera retrosynthesis model published in Nature, tackling drug discovery's synthesis bottleneck — marwinsegler · 2026-09-22
- Hazy Research: agents are replacing abstractions — CUDA DSLs are heading to retirement — ricklamers · 2026-09-22
- OXtalign: open-source crystal-packing similarity tool up to 200x faster than COMPACK — CatAstro_Piyush · 2026-09-22