Information-theoretic Bounded Rationality: Timing Laws and Cognitive Costs Emerge from Rational Bounds on Information Coding and Transmission

University of Minnesota Ph.D. dissertation. October 2019. Major: Psychology. Advisors: Paul Schrater, Maria Gini. 1 computer file (PDF); xi, 164 pages.Cognitive models are used to characterize and understand task performance in humans. Human behavior often deviates from predictions made by models that assume perfect rationality. Imposing constraints on cognitive resources, time, and/or information, while still assuming optimal function within those bounds, produces better characterizations of behavior. However, many of the proposed constraints and costs are ad-hoc and are not derived from fundamental limitations on computation. We suggest that behavioral performance is limited by the necessity of encoding and transmitting information about the world in the brain. Encoding information imposes a set of intrinsic bounds, defined by signal power, noise power, and knowledge of environmental statistics, that can be understood and quantified using concepts from information theory. In this dissertation, we investigate the patterns of behavior that should arise if cognition is subject to these bounds. Using an information transmission mechanism built using stochastic processes and Bayesian inference, we show that known `laws' of human behavior, including the Hick-Hyman law and the Power Law of Learning, are direct consequences of unavoidable limitations on the efficiency of information transmission. By instantiating constraints on information transmission in a working system, we are able to quantify transmission costs induced by task performance. This provides a unifying and principled explanation of cognitive costs and mental effort: effort arises in tasks that require expensive information transmission and is reduced through practice as learned task statistics are exploited to improve efficiency. To test the extent to which humans exploit task statistics to improve efficiency, we measured behavior on a version of the N-back task modified to include a predictable structure in target responses. We found that human data closely matches model predictions, suggesting that humans integrate information about both task structure and past images to produce responses. This finding is an experimental validation of our model, and suggests that the N-back task is more complex than is normally assumed. In sum, we show that treating cognition as a process constrained by fundamental bounds on information transmission provides a unified explanation of a wide range of behavioral phenomena