Computational and cerebral characterization of perceptual categorization mechanisms under uncertainty

Jury
Anne Urai (rapportrice, Leiden University)
Benedetto De Martino (rapporteur, University College London)
Alizée Lopez-Persem (examinatrice, Sorbonne Université)
Claire Sergent (examinatrice, Université de Paris)
Fabien Vinckier (examinateur, Université de Paris)
Valentin Wyart (directeur de thèse, Ecole normale supérieure)

ABSTRACT

Uncertainty represents an ubiquitous challenge for perceptual decision-making. Reducing uncertainty to make accurate decisions requires accumulating unreliable sensory information through Bayesian inference. Nonetheless, this cognitive process is inherently suboptimal in human observers, largely contributing to the substantial variability in human decisions under uncertainty. These ‘inferential errors’ are often underestimated by existing theories. Instead of ignoring those important limitations and approximating decision-making with optimal models, this thesis aims at better characterizing the cognitive strategies deployed by human observers to efficiently overcome these limitations. In the first experimental study of this thesis, we investigated human categorical decisions made in changing environments. In such environments, inferential errors interfere with accurate decision-making by triggering unwarranted changes-of-mind. Using computational modeling of behavior, we show that human observers compensate for this variability using a conditional belief updating strategy. Instead of accumulating each piece of noisy information available, they only update their beliefs when the incoming information is judged as sufficiently reliable. This parsimonious strategy not only reduces cognitive costs by preventing unwarranted changes-of-mind, but also increases decision accuracy. In the second experimental study of this thesis, we compared how human observers mitigate the costs of suboptimal inference across different forms of uncertainty that are used interchangeably in the literature: sensory uncertainty (triggered by noisy stimuli) and category uncertainty (triggered by ambiguous stimuli). We show that under category uncertainty, human observers do not deploy the conditional belief updating strategy observed in the first study and instead rely on their prior beliefs to stabilize their behavior. We further show in this second study that pupil dilation dynamics reflect this change in cognitive strategy across the two forms of uncertainty. Finally, in the third experimental study of this thesis, we investigated how information is represented during the inference process underlying categorical decisions. Existing theories propose that perceptual inference is either implemented in the ‘native’ sensory format that represents decision-relevant stimulus features, or in the ‘output’ format that encodes the action plan resulting from the decision. Instead, we provide behavioral and neuroimaging evidence that information is accumulated in an intermediate and abstract ‘category’ space defined by current decision alternatives. This ‘compression’ of stimulus information represents an efficient strategy to compensate for inferential errors as it allows to only accumulate the information that is relevant for the upcoming decision (and ignore decision-irrelevant dimensions). Using a carefully designed task which manipulates participants’ ability to perform information accumulation either in its native sensory or in a more abstract category space, we found that human observers preferentially accumulate information in category space. For the exact same amount of information, this strategy improves the accuracy of human decisions. Multivariate pattern analysis of magnetoencephalography (MEG) signals further validated our hypothesis as we could decode a neural signature of this compressed representation during information accumulation, whose precision was predictive of participants’ decision accuracy at the single trial level. Overall, this thesis underlines the flexibility of human cognition under uncertainty, and points toward future research.