Role of prefrontal cortex in rule and category learning: a neural network model

S. E. LEW, D.A. GUTNISKY, E. CYNOWIEC & B. S. ZANUTTO

New Orleans, USA

Congreso; 33th Annual Meeting of the Society for Neuroscience.; 2003

Society for Neuroscience

We present a neural network model to simulate the categorization of perceptually similar stimuli. Many authors (e.g. Miller & Cohen, 2001) found that the prefrontal cortex (PFC) plays an important role in rule learning. Main nervous structures simulated in the model are: the regions of the PFC involved in the learning of rules (PFC-L) and in the computation of short-term memories (traces); the ventro tegmental area and sustantia nigra (VTA-SN), as well as the premotor and the motor cortex (PMC-MC), and structures involved in the computation of the novelty (e.g. locus coeruleus). Inputs to the model are CSs and the US. Neurons in the VTA-SN learn to predict US updating their synapses by a modified Rescorla-Wagner rule. This prediction activates neurons in the PFC, increasing their probability of firing. Neurons in the PFC and the PMC-MC update their synapses by a hebbian or an anti-hebbian rule depending on the prediction of US. In our previous model (Lew & Zanutto, 2002) each neuron in the PFC-L represents a group of biological neurons involved in the same process. In this work we have replaced those individual neurons with a population of cells to study their emergent properties. After learning, a map of trained rules that emerges from the PFC plays a key role in the execution of responses according to these rules. The model learns DMTS performing close to 90% accuracy and categorization of perceptually similar stimuli. The model predicts that during the learning of the rules, neurons in the PFC-L are massively activated, but once the rules are learned, only a small percentage of them continue firing. It also predicts the results of Miller and Duncan’s experiments where two categories, dogs and cats, are learned. As in their experiments, we found neurons in PFC-L that codify categories (Miller 2001) and respond stronger to new learned stimuli than to older ones (Duncan 2000) when morphed stimuli are presented to the model.