UC Santa Barbara

Computational modeling is indispensable in the pursuit of understanding how the brain generates some of our most intimate subjective experiences, and how it solves some of the most interesting problems posed by our environment. The first model presented in this dissertation attempts to improve our understanding of how humans generate subjective liking judgements of stimuli, while two additional models are presented that attempt to improve our understanding of how humans learn adaptively and flexibly in a changing environment. This dissertation begins by introducing the theoretical foundations underlying these models and then proceeds by introducing each of them independently.

The first model is a probabilistic multidimensional model that accounts for both sensory and hedonic ratings collected from the same experiment. The model combines a general recognition theory model of the sensory ratings with Coombs' unfolding model of the hedonic ratings. The model uses sensory ratings to build a probabilistic multidimensional representation of the sensory experiences elicited by exposure to each stimulus, and it also builds a similar representation of the hypothetical ideal stimulus in this same space. It accounts for hedonic ratings by measuring differences between the presented stimulus and the imagined ideal on each rated sensory dimension. Therefore, it provides precise estimates of the sensory qualities of the ideal on all rated sensory dimensions. The model is tested successfully against data from a novel experiment.

The second model is a neurocomputational model of the flexible learning of abstract rules. The model is constructed from highly simplified building blocks that each represent a different brain region. It implements win-stay and lose-switch signals, and it computes and represents predicted rewards. Despite its simplicity, the model gives an impressively accurate qualitative and quantitative account of some challenging behavioral and neural data.

The third model uses a network of spiking neurons to represent activity within a neural circuit that implements adaptive learning rates by modulating the gain on the dopamine response to reward prediction errors. The model generates a dopamine signal that depends on the size of the tonically active dopamine neuron population and the phasic spike rate. The model was tested successfully against results from two single-neuron recording studies and a fast-scan cyclic voltammetry study. The general applicability of the model to dopamine-mediated tasks transcend the experimental phenomena it was initially designed to address.

This dissertation concludes with the most instrumental findings that surfaced through the process of creating the three models. It will give a behind the scenes view of the process of model invention and offer some practical advice for creating computational cognitive neuroscience models.