Periodic Reporting for period 1 - ACTOR (Towards a computational account of natural sequential behavior)
Période du rapport: 2022-10-01 au 2025-03-31
Imagine navigating through your kitchen, walking around your table and chairs, reaching the counter, sliding a cup on the countertop, pouring yourself some coffee, and preparing a sandwich. Such an everyday situation requires a sequence of behaviors involving the sophisticated interplay of perception, attention, cognition, planning, and action, all unfolding in an ambiguous and uncertain world. Yet, although past research in the behavioral sciences has revealed many properties of these constituent faculties, understanding how humans coordinate these cognitive abilities when carrying out everyday tasks is still rudimentary.
This project aims to conduct behavioral experiments and develop computational models to identify and characterize perception, cognition, and action in the wild. Our hypothesis is that sequential actions involving perceptual uncertainty, action variability, internal and external costs, and benefits can be understood and modeled in the unified framework of optimal sequential decision-making under uncertainty. To achieve this goal, we integrate ideas and methods from psychophysics, statistical decision theory, active vision, intuitive physics, optimal control under uncertainty, navigation, and inverse reinforcement learning. Our research considers tasks with increasing naturalness, from classical psychophysical production tasks and visuomotor control to navigation and sandwich making. We have developed methods that allow recovering subjects’ perceptual uncertainties, beliefs about tasks, and their subjective costs and benefits, including effort and action variability. Based on our previous experimental work on visuomotor behavior and theoretical work in modeling such behavior on an individual-by-individual and trial-by-trial basis, we seek to elucidate how perception, cognition, and action are inseparably intertwined in extended, sequential behavior.
This project aims to conduct behavioral experiments and develop computational models to identify and characterize perception, cognition, and action in the wild. Our hypothesis is that sequential actions involving perceptual uncertainty, action variability, internal and external costs, and benefits can be understood and modeled in the unified framework of optimal sequential decision-making under uncertainty. To achieve this goal, we integrate ideas and methods from psychophysics, statistical decision theory, active vision, intuitive physics, optimal control under uncertainty, navigation, and inverse reinforcement learning. Our research considers tasks with increasing naturalness, from classical psychophysical production tasks and visuomotor control to navigation and sandwich making. We have developed methods that allow recovering subjects’ perceptual uncertainties, beliefs about tasks, and their subjective costs and benefits, including effort and action variability. Based on our previous experimental work on visuomotor behavior and theoretical work in modeling such behavior on an individual-by-individual and trial-by-trial basis, we seek to elucidate how perception, cognition, and action are inseparably intertwined in extended, sequential behavior.
We have developed experiments together with mathematical tools to infer participants’ perceptual uncertainty, action variability, and internal cost functions from behavioral data. We formalize single decisions under uncertainty in a production task and thereby explain a broad range of successes but also specific and systematic biases in human behavior, e.g. pervasive undershoots in classical psychophysical tasks as the consequence of trade-offs resulting from perceptual uncertainty, action variability, error avoidance preferences, and the actor's cognitive effort. We have developed such a methodology, i.e. a doubly Bayesian model in which we utilize probabilistic programming together with amortized inference using neural networks to infer the latent properties of human behavior. This work can be seen as a form of rational analysis by asking what behavior has been optimal for. As such, this method reconciles normative and descriptive models of human behavior.
A few years back, a new experimental paradigm called continuous psychophysics was introduced to measure perceptual performance in humans. We have developed an analysis technique for such experiments in the framework of optimal control under uncertainty and solved the associated inverse problem so that estimated perceptual uncertainties match those obtained in corresponding classical psychophysics experiments better than previous analysis techniques. Importantly, the methods that we have developed allow inferring not only the perceptual uncertainty from such experiments but additionally allow recovering participants’ motor variability, subjective behavioral costs such as effort, as well as possibly false beliefs about the experimentental dynamics.
We originally wanted to develop a model of human navigation integrating self-motion and landmark cues involving perceptual uncertainty, action variability, and external as well as internal costs and benefits in the framework of simultaneous localization and mapping to elucidate the origins of known biases and variability in human trajectories and end-points in navigation experiments. We extended the scope of the modeling also to include the active planning of walking relative to an uncertain internal representation of space, leading to a model of belief-space planning under uncertainty. This work has elucidated the continuous, sequential, and dynamic nature of the active shaping of uncertainties in human navigation and is able to account for a wealth of previously published data.
A few years back, a new experimental paradigm called continuous psychophysics was introduced to measure perceptual performance in humans. We have developed an analysis technique for such experiments in the framework of optimal control under uncertainty and solved the associated inverse problem so that estimated perceptual uncertainties match those obtained in corresponding classical psychophysics experiments better than previous analysis techniques. Importantly, the methods that we have developed allow inferring not only the perceptual uncertainty from such experiments but additionally allow recovering participants’ motor variability, subjective behavioral costs such as effort, as well as possibly false beliefs about the experimentental dynamics.
We originally wanted to develop a model of human navigation integrating self-motion and landmark cues involving perceptual uncertainty, action variability, and external as well as internal costs and benefits in the framework of simultaneous localization and mapping to elucidate the origins of known biases and variability in human trajectories and end-points in navigation experiments. We extended the scope of the modeling also to include the active planning of walking relative to an uncertain internal representation of space, leading to a model of belief-space planning under uncertainty. This work has elucidated the continuous, sequential, and dynamic nature of the active shaping of uncertainties in human navigation and is able to account for a wealth of previously published data.
ACTOR has already significantly advanced our scientific understanding of perception and action in sequential sensorimotor behavior, exemplarily in human navigation and in the newly developed experimental paradigm of continuous psychophysics. We have developed a probabilistic model of navigation that integrates uncertainties in perception, cognition, and action. Importantly, as it includes the process of path planning relative to an uncertain internal map, it explains how humans actively shape the relative uncertainties about space over time. Because the uncertainties about the walker’s own position and the positions of the objects and landmarks are continuously and dynamically influenced by the path the walker plans and carries out, perception and action are intertwined. Accordingly, the model gives rise to active sensing strategies in which the reduction of uncertainty about the world is balanced with navigational goal achievement. This model can quantitatively reproduce the specific successes and errors and the different patterns of variability in walking trajectories across numerous triangle completion experiments from several laboratories across the world collected over 15 years with a single set of parameters and reconciles previous seemingly contradictory accounts, e.g. of cue integration in human navigation. A second major achievement is an analysis method of continuous psychophysics, which allows inferring participants’ perceptual uncertainty, internal subjective costs such as effort, possibly false beliefs about the experiment’s dynamics, and motor variability. This clarifies how perception, cognition, and action are intertwined in sequential behavior.