Abstract

Normative models of behavior strive to explain why behavior unfolds the way it does and have been highly successful in explaining many phenomena in neuroscience, cognitive science, and related fields. The power of these approaches derives from the combination of controlled experimental designs with their associated normative models, e.g. forced-choice psychophysics experiments with Bayesian observer models. Unfortunately, these tasks do not have much in common with real-world behavior, as they divide behavior into independent trials with discrete responses, often by highly trained participants. In naturalistic tasks, however, behavior is typically continuous and sequential. While highly controlled classical psychophysics tasks allow using normative models to estimate perceptual uncertainty and biases, naturalistic tasks introduce additional cognitive and motor factors such as action variability, intrinsic behavioral costs, and subjective internal models. To account for these factors, I propose to apply inverse normative modeling, i.e. to infer the components of normative models from behavior. In this talk, I will first present recent work that extends Bayesian models of perception to more general cost functions including intrinsic behavioral costs. I will then apply inverse normative modeling to continuous psychophysics. This recently developed experimental approach abandons the rigid trial structure of classical psychophysics and replaces it with a more naturalistic and intuitive continuous tracking task. It produces more temporally fine-grained measurements and allows efficient data collection even with untrained participants. Using Bayesian inverse optimal control, perceptual uncertainty, action variability, behavioral costs, and subjective beliefs about the task dynamics can be estimated from behavior in a tracking task. Finally, I will discuss some limitations of the method and show recent methodological extensions that address these limitations and allow applying inverse optimal control to a wider range of tasks. In summary, these methods open up the possibility of fitting normative models to more naturalistic continuous behavior.