Диссертации, представленные на защиту и подготовленные в НИУ ВШЭ

Recent studies showed that the visual system represents the whole distribution of objects’ features presented in a scene. Present work develops this idea and explores the ways how this rich representation can be used. First, we showed that the visual system can extract different ensemble summary statistics for a group of objects (the mean value, range, and numerosity) in parallel (without mutual interference) and using independent cognitive mechanisms. Secondly, the shape of feature distributions is used as an important cue for rapid visual categorization and segmentation of multiple objects defined by a conjunction of two features. Finally, we showed that the visual system strongly relies on the feature distribution representation during several types of visual search.

The dissertation presents a collection of seven articles on the general issue of the representational format underlying an ability to perceive multiple objects and to store them in working and long-term memory. Representing large sets of things is supposed to be a computationally demanding problem because of fundamental limitation of deep object processing. The basic framework presented in the dissertation suggests that the visual system can recover object information from sparse and compressed representations that do not necessary require full access to deep object processing (e.g. focused attention). Overall, three hierarchically organized types of representations are discussed: (1) features - separable object properties such as color or size, separate image parts, etc; (2) genuine whole objects, and (3) ensembles - global statistical summaries of multiple object features, such as average speed or size variance. Features are shown to play an important role in organizing visual experience: For example, complex everyday objects can be decomposed to independent semantically meaningful features in long-term memory (Utochkin & Brady, 2020a). On the other hand, ensemble representations can give easy access to quite rich representations of multiple objects without need to know everything about each individual object (Khvostov & Utochkin, 2019; Tiurina & Utochkin, 2019) and can be used in a number of visual tasks including categorization, visual search, texture discrimination (Utochkin, 2015; Utochkin & Yurevich, 2016; Utochkin, Khvostov, & Stakina, 2018), and in working memory encoding and storage (Utochkin & Brady, 2020b).