Stefano Fusi: The geometry of abstraction in prefrontal cortex and in the hippocampus

Columbia University

Abstract cognitive entities ("concepts") are constructed from multiple specific instances and enable generalization upon encountering new situations. One way of representing an abstract concept in a neuronal ensemble is to discard all information about specific instances. However, only a single concept can be represented in this way.  Here we show how a neuronal ensemble can represent multiple abstract variables simultaneously. We characterize the geometry of these representations, and we show that multiple abstract variables (context, action and value) are represented in dorsolateral pre-frontal cortex, anterior cingulate cortex, and the hippocampus of monkeys trained to perform a serial reversal learning task. The proposed geometry allows for generalization and it is observed also in a simul<div> </div>
ated multi-layer neural network trained with backpropagation. These findings provide a novel framework for understanding how the brain represents multiple abstract variables simultaneously and gives a new meaning to area specialization.

Organized by

John-Dylan Haynes/Margret Franke

Location

BCCN Berlin, lecture hall, Philippstr. 13 Haus 6, 10115 Berlin