Considering Reaction Time within an Accumulator Model of Decision-Making

Abstract

Animals take in and process sensory information in order to make decisions. The Brody lab at Princeton has administered an auditory decision task to rats, playing randomly-generated trains of "clicks" in each of their ears and prompting them to decide which ear heard more clicks [1]. An accumulator is an abstract variable that models how animals might acquire evidence towards making decisions. Brunton, Botvinick, and Brody added parameters to the accumulator to account for possible variations in accumulation, like inherent noise or bias towards one side. Through a "forwards-backwards pass", they fit the model on data, obtaining max likelihood parameters for use in explaining animal decision-making behavior. In this thesis we make two extensions to the Brunton accumulator model: first, we reformulate the bias parameter from a bias present at the time of decision to a bias present before evidence occurred. This necessitates an updated gradient calculation which is verified empirically. Fitting rat data on both types of bias simultaneously reveals that the two formulations are basically interchangeable under otherwise optimal parameters. Next, we modify the model to accept reaction time information from rats, allowing them to decide before being prompted. Although reaction time data has not yet been gathered on the auditory click task, we generate artificial rat decision times from real rat data in order to test the implementation. We first use this to show the effect of parameter variation on reaction time. Lastly, we fit the artificial data and show that we can indeed retrieve generating parameters as maximum likelihood parameters. 2