Reconstruction of cue-invariant 3D appearances from brain activity by Shuntaro C. Aoki, Reo Tsukasa, Shiyun Yang, Misato Tanaka, Eizaburo Doi, Tomohiro Nakamura, Jun-Kai Ho, and Yukiyasu Kamitani has been posted on bioRxiv. The study shows that a cue-invariant internal representation of 3D structure can be externalized from human fMRI activity as explicit 3D objects: fMRI responses are decoded into the latent features of a pretrained 3D point-cloud autoencoder, and a generator maps them back to a 3D shape. A decoder trained only on responses to 2D rendered images generalizes across depth cues to random dot stereograms (RDSs)—which convey 3D solely through binocular disparity—and reconstructs the slant of contour-matched RDSs whose 2D outlines are held identical, with the cue-invariant signal strongest along the dorsal visual stream. [bioRxiv] [Rotating reconstructions — natural objects] [Rotating reconstructions — artificial objects (2D & RDS)] [Data (OpenNeuro)]