Generally, my research focuses on using computational models to understand how cognition and behavior are supported by brain systems, like the cerebral cortex, basal ganglia and hippocampus. In addition to mathematical models, I have also used neural network models like feedforward neural networks, recurrent neural networks of spiking neurons (STDP), simple recurrent networks (backprop). One aspect of my research involves using using mathematical models to study how multiple cognitive systems interact to give rise to intelligent behavior. For example, computational models of eye movement control during reading describe how various perceptual, cognitive and motor processes interact to result in patterns of eye movements.

Often, computational models raise questions (or in some cases, generate "predictions") that need to be tested experimentally before further computational work can be done. In attempting to answer such questions, I have obtained experience in eye-tracking (e.g., gaze-contingent paradigms) and fMRI (e.g., fast-event related designs and model-based fMRI analysis).

Modeling Philosophy

From a theoretical perspective, computational models are useful for understanding how principles of neural circuits can explain cognition and behavior. One way they do this is that they make it easy for investigators to make changes to the assumptions used in building the model, effectively allowing investigators to perform highly-controlled experiments.

From a practical perspective, implemented models can be used to make testable predictions about behavior and/or brain activity. If the model makes correct predictions in many circumstances, it is likely to be correct.  For example, one interesting application of computational models is to use their predictions to fit behavioral data and analyze fMRI data (i.e., model-based analysis).

Two important points about models: (1) models should be simple enough for clarity and communicability; and, (2) knowledge from neuroscience provides additional constraints that make models more likely to be correct.   The latter point is important because radically different models could generate the same behavior in a given experiment.