Brain-IT: AI Decodes Vision in One Hour

Overview

Researchers at the Weizmann Institute of Science have unveiled Brain-IT, a groundbreaking artificial intelligence model capable of decoding visual perception from brain activity with unprecedented speed and accuracy. Unlike previous systems that require dozens of hours of individualized training, Brain-IT can reconstruct complex scenes after just one hour of scan data.

Published on September 14, 2026, this development marks a significant leap in the field of computational neuroscience. The model, led by Professor Michal Irani, identifies shared neural patterns across different brains, allowing it to "read" what a person is seeing with striking fidelity.

Breaking the Personalization Bottleneck

Recent advancements in neural interfaces have enabled scientists to restore speech to individuals with severe paralysis. However, these existing technologies rely on personalized AI models trained on tens of hours of brain recordings specific to each patient. While life-changing for users, these systems are far from universal mind-reading tools because they cannot generalize across different individuals without extensive retraining.

Brain-IT circumvents this limitation by focusing on universal neural patterns. The model recognizes activity sequences that recur across various brain regions in different people as they process visual information. This approach allows the AI to adapt to a new user rapidly, requiring only sixty minutes of fMRI data to achieve high-quality reconstruction.

“There exist nowadays models that translate brain activity into images... However, they tend to make mistakes in basic features such as composition and color. The new model we developed outperforms them in reconstructing both the content of the image and its details,” explains Prof. Michal Irani.

A Novel Training Strategy for Data Scarcity

The path to creating Brain-IT was hindered by a critical shortage of high-quality training data. Generating the necessary datasets typically involves placing participants inside claustrophobic MRI machines for extended periods while they view thousands of images. Consequently, the largest existing datasets in the field contain scans from no more than eight individuals—a mere fraction of what is needed to train robust deep learning algorithms.

To overcome this scarcity, the Weizmann team devised a solution analogous to a two-way bilingual dictionary. They developed not only a decoder that reconstructs images from fMRI scans but also an encoder that predicts brain activity from images.

During the training phase, the encoder processes random images that have never been viewed by a human in an MRI scanner. It generates estimated brain scans for these images, which are then fed into the decoder. The decoder attempts to reconstruct the original image from these synthetic scans. Through this iterative back-and-forth translation, the models effectively build their own massive, self-generated dataset, learning to produce scans that encode visual information effectively without requiring every single scan to be physically performed.

Discovering Universal Functional Regions

The primary technical challenge in building the encoder was accounting for the natural variation between individual brains. The same image can evoke distinctly different patterns of neural activity depending on the person. To address this, the researchers segmented the problem at a granular level.

Instead of attempting to encode an entire image into a whole-brain scan at once, the team divided the brain into approximately 40,000 brain voxels—tiny three-dimensional units of volume—and measured how each responded to specific visual stimuli. They combined this with an AI model that extracted both basic visual features, such as color and location, and semantic features, such as the presence of a face or a food item.

This multi-layered analysis revealed recurring patterns of brain activity associated with specific image features, enabling the creation of a universal brain encoder. Remarkably, during this process, the encoder naturally identified 128 functional regions shared by all humans that play specific roles in image processing.

While some of these regions were already known to neuroscientists, others appear to be entirely new. For instance, the study highlighted a division of labor within the brain region responsible for processing scenes and places, offering fresh insights into human visual cognition.

Performance and Recognition

The capabilities of Brain-IT were demonstrated through high-fidelity reconstructions of complex scenes. In one notable example, the model successfully reconstructed a baseball game viewed by a participant, capturing both the semantic content and intricate details such as the position of the batter and the ball in the air.

The results indicate that Brain-IT preserves not only the general meaning of an image but also its compositional structure and color palette, areas where previous models often faltered. This dual improvement in speed and accuracy positions Brain-IT as a significant advancement over prior state-of-the-art methods.

Due to its innovative methodology and potential impact, the research was selected for presentation at the International Conference on Learning Representations (ICLR). ICLR is recognized as one of the world’s leading conferences in the field of artificial intelligence, underscoring the scientific community's interest in this breakthrough.

The development team, working within the Department of Computer Science and Applied Mathematics at the Weizmann Institute, includes Roman Beliy, Amit Zalcher, Jonathan Kogman, and Navve Wasserman. Their work bridges the gap between computational efficiency and neurological accuracy, bringing the vision of rapid, non-invasive brain imaging one step closer to clinical and practical application.