Abstract Neurons of the medial temporal lobe (MTL) form the basis of semantic representation in the human brain. While known to contain category-selective cells, it is unclear how the MTL processes naturalistic, dynamic stimuli. We studied 2286 neurons recorded from the hippocampus, parahippocampal cortex, amygdala, and entorhinal cortex of 29 intracranially-implanted patients during a full-length movie. While few neurons responded preferentially to semantic features, we could reliably predict the presence of characters, location, and visual transitions from the neuronal populations using a recurrent neural network. We show that decoding performance differs across regions based on the feature category, and that the performance is driven by feature-selective single neurons when decoding visual transitions such as camera cuts. These findings suggest that semantic representation in the MTL varies based on semantic category, with decoding information embedded in specific subsets of neurons for event-related features or distributed across the entire population for character and location-related features.

Decoding movie content from neuronal population activity in the human medial temporal lobe