Abstract: Breaking down difficult problems into simpler parts and, conversely, composing primitive units to generate rich behavior, are hallmarks of biological intelligence. Principles of reductionism and compositionality may also guide how we process the complex visual environment of our everyday experience—that is, natural visual processing. This talk will be centered on a core element of this strategy, broadly termed grouping and segmentation, by which our visual system organizes complex visual inputs into groups corresponding to distinct perceptual objects. I will present progress made over the years by my lab and collaborators through a tight integration of theory, computational modeling, visual neurophysiology and psychophysics.

First, I will present a normative framework that unifies two widespread observations in primary visual cortex (V1): spatial contextual modulation (how the activity of a neuron in response to a target stimulus is modulated by contextual stimuli) and response variability (fluctuations in neural activity across repeated presentations of the same stimulus). Specifically, we hypothesize that the computational goal of V1 is to approximate a probabilistic representation optimized to the statistics of natural visual inputs, and that the structure of V1 activity is best understood in the light of this goal. I will present a concrete computational framework that instantiates this hypothesis and reproduces a wide array of classical observations on contextual modulation and shared variability.

Second, building on that foundation, I will argue that a complete understanding of this phenomenology must also account for the non-stationary statistics of natural inputs. The theory makes detailed predictions about the sensitivity of V1 neurons to segmentation and grouping cues, including a surprising flexibility of functional interactions that we have confirmed recently with data recorded by our collaborators using multielectrode arrays in macaque V1.

The third piece will focus on perceptual grouping and segmentation in human observers. Extending our computational framework to deep probabilistic algorithms for natural image and video segmentation, leads to a surprising prediction about the time course of perceptual segmentation: processing time to judge if two parts of an image are in the same segment, increases with distance if they are in the same segment, but decreases with distance otherwise. I will provide empirical evidence recorded in a new experimental paradigm we have developed to measure perceptual grouping of natural stimuli with human participants. This effect challenges popular theories of the time-course of perceptual organization based. In our model, it results from the interactions between spatial priors and dynamic Bayesian inference, offering a normative foundation for recent semi-mechanistic models based on artificial RNNs, and a path to better align them with human perception.

Please, be informed that video/photo recording might take place at the event and the edited version of the video material might be published to communicate or promote CEU PU's activities. Please, find our Privacy Notice here.

Ildikó Varga

Department Coordinator (Budapest)

Department of Cognitive Science

H-1051 Budapest

Nádor u. 15. FT room 404.

tel: +36-1 327-3000 2941

http://www.ceu.edu

http://cognitivescience.ceu.edu

From: Talks <talks-bounces@cogsci.ceu.edu> on behalf of Ildiko Zsoka Varga <VargaI@ceu.edu>
Sent: Tuesday, August 27, 2024 1:02 PM
To: 'talks@cogsci.ceu.edu (talks@cogsci.ceu.edu)' <talks@cogsci.ceu.edu>
Subject: [CEU Cogsci Talks] CCC Colloquium - Ruben Coen-Cagli, Tuesday, Sept. 3. Budapest