A new artificial intelligence model reconstructs visual images from human brain scans and predicts neural responses based on visual inputs.

WASHINGTON — Researchers have developed an artificial intelligence system capable of reconstructing visual images directly from human brain scans with high precision, while simultaneously predicting neural activity based on visual stimuli.

The bi-directional model operates in two distinct modes. In the primary decoding function, the system analyzes brain scan data collected while a person views a specific image, translating the recorded neural patterns back into a reconstructed visual output. In the reverse encoding function, the system receives a visual image and predicts the corresponding neural response that a brain scan would record.

Bidirectional Neural Decoding

The technology relies on machine learning architecture designed to map relationships between complex neuroimaging data and visual features. Functional neuroimaging records localized activity across different regions of the human brain, generating spatial data when a subject views an object or scene.

By training on paired datasets of visual inputs and corresponding brain recordings, the algorithm identifies statistical connections linking visual properties—such as structural layout, shapes, and color distributions—to localized neural signals. When presented with previously unseen brain scan data, the system decodes those neural signals to generate a matching visual output.

Reconstruction Precision and Applications

Testing demonstrates that the visual reconstructions closely resemble the original target images viewed by human subjects during scanning sessions. The capability to run the system in reverse allows researchers to simulate expected neural responses without requiring continuous physical brain imaging.

Researchers indicate that bi-directional decoding models advance fundamental neuroscience by providing quantifiable methods to test theories of human visual perception. The technique also offers potential structural frameworks for non-invasive brain-computer interfaces, assistive communication systems, and neuro-prosthetic technologies.

Further refinement of the technology remains focused on improving signal resolution and increasing the structural fidelity of reconstructed images.