NB & CNS Seminar - Terry Sejnowski - Tuesday, January 31, 2023 - 4 pm
NB & CNS Seminar
Terry Sejnowski, Professor and Laboratory Head
Salk Institute for Biological Studies
Tuesday, January 31, 2023 - 4 pm
Title: Is the Hippocampus a Predictive Autoencoder?
Sequences of sensory experiences create long-term memories when the hippocampus is engaged. The hippocampus encodes these sequences and replays them during sleep to consolidate them in the cerebral cortex. The hippocampus receives inputs primarily from the cortex and its major output is back to the cortex, forming a loop. This anatomy has the form of an autoencoder, an artificial neural network model that reproduces inputs on the output layer after passing through an intermediate layer. This hidden layer learns to compressively encode the latent variables of the input distribution. The same architecture can learn sequences of inputs when the hidden layer is a recurrent network and the loss function is generalized to predict the next input. We show how this can be accomplished with biologically plausible local learning algorithms. The model reproduces a wide range of experimental observations including place cells and phase precession and makes predictions for the relative timing of neural activity in different parts of the hippocampus.