Weizmann Institute of Science’s (魏茨曼科学研究所) Brain-IT learns new-subject image decoding from approximately one hour of functional magnetic resonance imaging data.

The Weizmann Institute of Science (魏茨曼科学研究所) announced Brain-IT, an AI model that reconstructs image content from functional magnetic resonance imaging (fMRI) recorded while people view images. The described task is decoding responses to viewed images, rather than unrestricted access to a person’s thoughts.

According to the institute’s account, Brain-IT requires approximately one hour of data to learn decoding for a new participant, compared with the tens of hours typically needed by existing methods. The team also reported improved reconstruction of image content, composition, and color, although the supplied message provides no numerical accuracy scores.

The researchers developed an image encoder and decoder and used artificially generated predictions of brain activity to address limited training samples. During the research, the encoder identified 128 image-processing functional regions shared across people, some of which the team described as previously undiscovered.

The team said the research had been accepted at the International Conference on Learning Representations (ICLR). It is exploring sound decoding and possible video decoding, but the supplied report does not establish demonstrated results for either extension.