From fMRI to images: what Brain-IT reconstructs-and what it does not
Can AI read minds? Brain-IT, published at ICLR 2026, studies a more specific task: reconstructing viewed images from fMRI measurements under controlled experimental conditions.
What the model learns
Training pairs viewed images with the brain scans recorded while people view them.
fMRI indirectly reflects blood-oxygen changes related to brain activity. The model learns relationships between these measurements and image features.
How reconstruction works
In a simplified explanation, a new brain scan is mapped to structural and semantic clues: the rough layout of a scene and what may be in it.
These clues guide a generative model to reconstruct an image.
Why the errors matter
The baseball example in the video recovers a similar person and field, while red clothing becomes purple.
Reconstruction can recover partial information; it is not a lossless recording of perception. This experiment did not test secrets, memories, or dreams.
Watch the 81-second explainer
The full video has Mandarin narration and Chinese subtitles. Watch the complete video on Bilibili.
Sources
- Roman Beliy et al., Brain-IT: Image Reconstruction from fMRI via Brain-Interaction Transformer, ICLR 2026.
- Author project and examples
- Official proceedings
The mechanism in the explainer is simplified for a general audience.
Comments
No comments yet. Start the discussion.