Disentangled latent variables in the hippocampus
Ok, that is my short summary interpretation (for AI nerds) of an intriguing article by a Cedars-Sinai-led team. From the abstract:
"Here we characterized the representational geometry of populations of neurons (single units) recorded in the hippocampus, amygdala, medial frontal cortex and ventral temporal cortex of neurosurgical patients performing an
inferential reasoning task. We found that only the neural representations formed in the hippocampus simultaneously encode several task variables in an abstract, or disentangled, format. This representational geometry is uniquely observed after patients learn to perform inference, and consists of disentangled directly observable and discovered latent task variables."
Translation: the researchers recorded individual neurons in multiple brain areas (in patient volunteers undergoing surgical treatment for epilepsy) and were able to show that only in the hippocampus did neuronal activation patterns represent the same concept in different contexts, such as relationships between items. One canonical AI example from the early days of word2vec would be the ability to recognize king:man/queen:woman as a single concept.
Another intriguing result: participants who did not initially recognize a concept but were then told about it had their hippocampal neuronal activation patterns reconfigured quickly to accommodate the new insight -with one "example". Rapid learning of the concept lead to the same abstract representations as slow learning from examples.
See also link for a nice perspective on the paper.