Weizmann Institute develops Brain-IT AI to reconstruct images from brain activity

Researchers at the Weizmann Institute of Science have developed Brain-IT, an artificial intelligence model capable of reconstructing images a person is perceiving by decoding brain activity. Unlike previous models that require dozens of hours of brain scans to adapt to a new individual, Brain-IT can learn to “read” a new person in approximately one hour.

The system works by identifying shared patterns of brain activity across different individuals. By analyzing approximately 40,000 voxels, the team identified 128 functional regions shared between people. The model utilizes a dual approach: a decoder that transforms brain activity into images, and an encoder that predicts brain responses from images, allowing for the generation of synthetic data to improve training.

Michal Irani, a researcher at the institute, noted that while existing models can preserve semantic meaning, they often fail at basic visual features. Brain-IT aims to overcome these limitations by improving the reconstruction of image composition, object positioning, and color. The research has been selected for presentation at the International Conference on Learning Representations (ICLR).