Séminaire de Jean-Pierre Nadal

In this talk I will present a set of results on the modelling of the neural bases of perceptual decision making. In a typical experiment in humans and other animals, the task consists in making a simple decision, to categorize a stimulus (e.g. an image) into one of two categories (e.g. 'cat' vs 'dog'). Categorisation, or classification, is indeed a fundamental cognitive ability. A well-known perceptual consequence of categorisation, "categorical perception", is notably characterised by intra-categorical compression and inter-categorical separation: two items, close in stimulus space, are perceived closer if they belong to the same category than if they belong to different categories. Making use of information theoretic tools, I will show that such properties are necessary outcomes of the optimal neural encoding of the stimuli, whenever this coding is done in view of the categorization task. However, other properties can hardly be cast within an optimal inferential viewpoint. In experiments, stimuli are presented in sequences of trials. Analysis of the percentage of correct responses and of the reaction times reveals various types of correlations between consecutive trials - e.g. average reaction times longer for trials following one with an incorrect response, even when one is not told that there was an error. By considering biophysical models of recurrent neural networks, I will show that these effects can be understood as artefacts of the nonlinear neural dynamics. In addition, I will present the results of an experiment, carried out in collaboration with the Cognitive Department of the ENS (Paris), showing that these models can account for the confidence in one's decision. This talk is mainly based on joint works with Kevin Berlemont (NYU, Center for Neural Science) and Laurent Bonnasse-Gahot (EHESS, CAMS).