Séminaire de Yinan Cao

Vendredi 20 juin à 11h, Salle Henri Gastaut, INT

Yinan Cao (Laboratoire de Neurosciences Cognitives, Ecole normale Supérieure, Université PSL)
invité par Andrea Brovelli

Flexible human information seeking in decisions under uncertainty

Abstract: Human information-seeking often deviates from normative models, yet these deviations may reflect adaptive strategies that accommodate cognitive limitations and enhance the precision of decisions. In this talk, I will present two studies showing how seemingly suboptimal information-seeking policies can support accurate decisions under uncertainty. In the first study, we used a reward-free information-sampling task to examine exploratory choices driven by a desire for knowledge rather than immediate gain. Participants adopted a twophase strategy: repetitive sampling of each option to form initial beliefs, followed by targeted sampling of the most uncertain one, rather than uncertainty-guided sampling from the outset. While statistically suboptimal, this strategy appears to reduce switch costs and stabilize neural representations of evidence. Neural decoding revealed that belief-confirming evidence accumulated across repeated choices, with switches occurring only after a satisficing threshold was reached. In the second study, we used magnetoencephalography to track covertattention during multi-alternative decisions. Participants viewed arrays composed of three Gabor patches, each with distinct contrast. On different trials, they were instructed to choose the stimulus with either the highest or lowest contrast. The same physical stimuli thus mapped onto orthogonal decision values depending on the task frame. Using an inverted-encoding model trained on an independent retinotopic localizer, we reconstructed the time course of the expression of attention, reflected in stimulus-independent modulations in the strength of stimulus representations across visual retinotopic maps. This neural reconstruction revealed a two-stage process: an early covert deflection of attention to the least valuable option (reflecting an elimination strategy), followed by a shift of attention toward the most valuable option. Together, these findings suggest that what appears suboptimal on the surface may reflect structured computational principles, offering new perspectives on the structure of human cognition and informing models of intelligent behavior in both biological and artificial systems.