3 days ago
Israeli Scientists Develop AI to Reconstruct Images From Brain Scans
Scientists have made an AI system called Brain-IT that tries to turn brain scans into pictures.
The scans show patterns of activity in people’s brains while they look at images.
The team trained the system using scans from eight people.
To help it learn more without scanning many more people, they made another part of the system predict how an unseen image might affect the brain.
Brain-IT then uses brain activity patterns to rebuild an image.
The researchers say it can make results comparable to some methods that used much more recording time.
They also found 128 brain regions shared across people that help process images.
In the future, this kind of technology might help researchers understand vision or support communication, and the team is exploring whether it could work with sound.
Brain-IT was developed by Israeli scientists to reconstruct images from brain activity measured with fMRI scans.
The researchers used scans from eight people who each viewed several thousand images.
An encoder predicts brain activity for images not shown during a scan, providing additional training examples for the decoder.
The researchers say Brain-IT’s results are comparable to methods that previously used about 40 hours of recordings, while requiring one hour of data from a new subject.
The team identified 128 shared functional brain regions and is exploring possible uses in visual research, assistive communication, and decoding sound.
- Who
- Israeli scientists developed Brain-IT.
- What
- An AI system that reconstructs images from brain activity measured with fMRI scans.
- Where
- When
- Why
- To decode visual information from brain activity and potentially support research into vision and assistive communication.
Key facts
- System
- Brain-IT
- Training dataset
- fMRI scans from eight people, each of whom viewed several thousand images
- Additional training approach
- An encoder predicts brain activity for images not viewed in a scanner, generating additional training examples.
- Reported performance comparison
- The paper says results are comparable to methods that previously required around 40 hours of recordings.
- Data from a new subject
- The system reportedly achieved comparable results using one hour of data.
- Functional brain regions
- The encoder identified 128 regions shared across people.
- Possible future exploration
- The team is exploring whether similar methods could decode sound.
Sources
Related news
IIT Madras Maps Human Brain Through Imaging and Engineering
AI Helps Archaeologists Find Patterns in Ancient Evidence
AI Helps Astronomers Create Video of Black Hole Jet
Scientists Develop SORDINO Technique to Make MRI Scans Quieter
Four ChatGPT Prompts Turn Complex Topics Into Visual Diagrams
OpenAI Plans Visual Ads During ChatGPT Image Creation
FirstCry Shares Hit 52-Week Low
