Developing artificial neural network tools for cognitive modeling

点击翻译按钮获取中文摘要

Project Summary/Abstract Mathematical modeling is an essential tool to study the brain, behavior and cognition. Computational cognitive models state with simple mathematical equations how the brain may be manipulating information that supports how humans and animals interpret the world around them, make choices and adapt to new environments or events. Researchers can use computational cognitive models to quantitatively test the theories embedded in the models, by comparing model predictions with behavioral and neural data. Models also often have meaningful parameters that can be tuned to reflect how specific information is used, for example how much participants weigh prospective gains vs. losses in decisions, how willing participants are to explore new information vs. exploit the information they already have, or how confident they need to be before committing to a decision. In the context of psychiatric and neuro-degenerative diseases, computational modelers can ask whether models fit patients' behavior/neural activity differently than healthy controls, thereby explaining impairments as a difference in information processing; or whether they exhibit different parameters for the same models, showing different weighing of information. Thus, computational modeling provides important quantitative tools to understand how brain disease impacts behavior and cognition. However, such research requires statistical tools to quantitatively relate models to data - that is, to identify which models and which parameters explain the data best. Existing tools mostly rely on computing the likelihood of the data under the model, and are very powerful for a specific class of models. However, they leave out a much broader class of models for which the likelihood is too complex to compute. This class of models includes many simple and relevant models that embody reasonable theories of cognition, but these models are currently unexplored, because researchers lack the tools required to relate them to data. The goal of this proposal is to develop new tools for this class of models, using modern supervised machine learning techniques (with deep neural networks) that bypass the need to compute the likelihood, but do not require advanced expertise in ap- plied mathematics and are broadly generalizable to the whole class of models that are currently inaccessible to existing techniques. Specifically, we will develop tools to 1) identify which of multiple models explain a partici- pant's data better, 2) identify the value of model parameters that best explain a participant's data, and 3) infer how the model variables generated the participant's behavior, enabling us to relate these variables to brain data. This research will vastly increase the potential reach of computational techniques in neuroscience, enabling researchers to consider theories that are currently discarded for lack of tools. This is an important step toward broadening our understanding of mental illness and brain diseases.