Developing models of thalamocortical unsupervised attentional selection and competitive learning (360G-Wellcome-091998_Z_10_Z)

The project will aim to establish the following. (1) The characteristics of a modular neural network architecture for unsupervised learning in which each module receives weighted predictions (expectations) and bottom-up activation, calculating a 'performance' output and a mismatch error. (2) The impact of introducing global competition for learning based on locally derived prediction error. (3) The subsequent effects on learning rates of introducing biologically based synaptic autostabilization mechanisms. (4) The networks' performance in simple learning-theory paradigms and in perceptual identification with prior information. (5) The unification of such models with methods for implementing high-speed dynamic binding. (6) The effects of network-wide parameter manipulations, including manipulations relevant for psychopathological states such as hallucinations. (7) The simulation of a large-scale version of the computational network and analysis of its topological and statistical properties, for comparison with existing MRI analyses. (8) The simulation of a three-dimensional version of the network and the derivation of predictions relating to surprise in learning, for comparison of the model's predictions with MEG experiments in normal humans and then patient groups.