Proceedings of the Annual Meeting of the Cognitive Science Society

It has been proposed that the basal ganglia is involved in selecting which action is performed, while the motor cortex specifies how the selected action is carried out. However, recent electrophysiological evidence from a reach-to-pull task challenges this view by showing that the motor cortex and the dorsal striatum in the basal ganglia jointly specify continuous action movement parameters, such as reach angle and pull force. This finding implicates the basal ganglia in both action selection and action specification, creating a gap for a mechanistic model that unifies both functions in a single, grounded framework. We address this gap with a biologically plausible model of the basal ganglia that leverages dynamic neural fields applied to high-dimensional representations of action-salience distributions. The dynamic neural field converts high-entropy representations into low-entropy representations without altering the underlying action-salience distribution's peak---a property previously shown to support accurate readout of continuous action parameters (e.g., movement speed or force). Our model is an extension that applies the same dynamic neural field transformation across multiple actions so that we can simultaneously (i) resolve which action is chosen (e.g., lever pull vs. button press) and (ii) specify the continuous action parameter value with which the chosen action is to be executed (e.g., force or speed). In our simulations patterned on the reach-to-pull paradigm, the basal ganglia model performs both discrete action selection and action specification of a continuous action parameter in a single operation, offering a concise mechanism for how basal ganglia-cortex circuits decide what to do and how to do it.