Overview
- Researchers from the Weizmann Institute presented Brain‑IT at ICLR 2026 on Tuesday, reporting that the model can reconstruct static images a person is viewing from their fMRI data.
- Brain‑IT groups thousands of fMRI voxels into about 128 functional regions that are common across people and uses those shared patterns to map brain activity back to visual features.
- Because the model learns cross‑subject patterns, it needs roughly one hour of recordings to adapt to a new participant instead of the many hours required by earlier methods.
- The team notes clear technical limits: fMRI measures blood‑oxygen changes roughly every two seconds, which prevents reliable decoding of fast video, dreams, or internal thoughts and keeps the work laboratory bound.
- Researchers plan to extend the method to sounds and video and to explore assistive brain–computer interfaces, and they emphasize ethical and privacy safeguards before any real‑world use.