Interpretable abstractions of artificial neural networks predict behavior and neural activity during human information gathering

Abstract

It has been suggested that humans and other animals are driven by a fundamental desire to acquire information about opportunities available in their environments. Not only might such a desire explain pathological behaviors, but it may be needed to account for how everyday decisions are resolved. Here, we combine artificial neural networks (ANNs) with symbolic regression to extract an expressive yet interpretable model that specifies how human participants evaluate decision-relevant information during choice. This model accounts for behavior in our own data and in previous work, outperforming existing accounts of information sampling such as the Upper Confidence Bound heuristic. This modelling approach has broad potential for uncovering novel patterns in behavior and cognitive processes, while also specifying them in human-interpretable formats. We then used the value of information derived by our model, together with ultra-high field neuroimaging, to examine activity across a suite of subcortical neuromodulatory nuclei and two cortical regions that influence these nuclei. This established roles for midbrain dopaminergic nuclei, anterior cingulate cortex, and anterior insula in mediating the influence of value of information on behavior.