CompCog: Bridging Levels of Analysis: Characterizing Algorithmic Models by Extreme Bayesian Priors

点击翻译按钮获取中文摘要

The field of cognitive modeling seeks to understand human thought and behavior using the languages of mathematics, statistics, and computers. A cognitive model is a set of equations or a computer program that can mimic how people act in experimental or real-world settings. These models are useful in many ways. They can predict how people or groups will act in new situations. They can guide development of educational materials and training systems that maximize learning. They can give insight into the inner workings of the mind, which can contribute to treatment of psychological and brain disorders. They can help to explain human intelligence and creativity, leading to new methods in artificial intelligence and machine learning. This project aims to further these goals through mathematical advances and human behavioral experiments that together may lead to new, more accurate models. A variety of interdisciplinary collaborations and outreach efforts will then explore application of these models to improving psychiatric diagnosis, developing new analysis methods for neuroimaging data, making artificial intelligence more comprehensible to the user (explainable AI), and making psychological models, statistics, and AI more accessible to undergraduate and high school students.The technical portion of this project investigates connections between two types of cognitive models: algorithmic and rational. Algorithmic models describe the mind in terms of information processing, specifying mental representations and the processes that act on them in going from perceptual input to observed behavior. Rational models explain a person’s learning and decision making in terms of his or her goals and beliefs about the how the world works. They assume the mind is highly tuned to its environment, and thus that it acts optimally relative to the inherent uncertainty in the world. Researchers usually think of algorithmic models as heuristics (i.e., simplified shortcuts) that approximate rational ones. Under this interpretation of algorithmic models, cognition falls short of being optimal because of physical limitations of the brain, such as how much it can remember or how much information it can process at once. This project will develop a different connection. Using formal mathematical analysis, the investigators will show how influential algorithmic models in psychology exactly match certain rational models under the assumption that the world is extremely uncertain and unpredictable. This connection will be used in several ways to develop new models: more sophisticated rational versions of existing algorithmic ones, more efficient algorithmic versions of existing rational ones, and intermediate models that combine the strengths of rational and algorithmic ones. Four series of experiments, each spanning tasks of decision making, reward learning, and concept acquisition, will test which models best predict human behavior, and also which yield the best objective performance in natural settings. If successful, the project will yield new mathematical foundations for the field of cognitive modeling, specific models that more accurately match human behavior, new tools for AI and statistics, and a new perspective on foundational questions of rationality of the human mind.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.