Conference on Cognitive Computational Neuroscience

Compared to all other animals, humans exhibit a remarkable ability to respond flexibly. For example, after even brief instructions, we can select appropriate responses despite no previous experience of the task (zero-shot learning). Most existing cognitive models of this are non-neural or non-quantitative. We recently proposed a class of models that simulates one plausible way that neurons might perform this computation. In our untrained, Hopfield-like artificial neural network, fast Hebbian plasticity rapidly creates and destroys new attractor states. This creates associations in a readily accessible buffer, that guide neural activity. Such a mechanism might allow human prefrontal cortex to flexibly switch between response mappings. This small dynamic neural model has high explanatory power. We examine a unique published dataset where humans received instructions to use one of two stimulus dimensions to make decisions. Using fMRI, both dimensions could be decoded. After TMS to frontal lobe, decoding of the relevant dimension was reduced. We simulate a transient perturbation in the network and reproduce this effect. It arises because when features become relevant, they rapidly form stronger synapses to frontal areas, and are thus more corrupted by the brain stimulation. This suggests rapid synaptic changes could explain flexible prefrontal control of action.