Geoffrey Hinton
University of Toronto

"What are lateral connections for?"

I shall start by describing a novel unsupervised learning algorithm and then I shall show how it provides an unexpected interpretation of the role of lateral connections in cortex. Recurrent lateral interactions between cortical cells are typically viewed as a way of enhancing the initial representation produced by the feed-forward pathway. For the algorithm I shall describe the role of the lateral interactions is to over-simplify the initial representation by removing details that do not fit with the brain's model of the world. By comparing the correct, initial representation with its over-simplified version the brain can learn to improve its model. By using synapses with the right temporal properties, later stages of processing can then associate responses with the fleeting peculiar details of specific episodes, thus avoiding interference with other associations.