Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

Iterative Bayesian inference explains the dynamics of perceptual organization of natural scenes

Tridib Kalyan Biswas¹, Jonathan Vacher², Pascal Mamassian³, Sophie Mollholm, Ruben Coen-Cagli⁴; ¹Einsteinmed, ²Université Paris-Cité, ³Ecole Normale Supérieure de Paris, ⁴Albert Einstein College of Medicine

Presenter: Tridib Kalyan Biswas

Perceptual decision making is classically conceptualized as evidence integration theory – the notion that sensory inputs are perceived by sequentially accumulating noisy samples from the environment and averaging out the noise. Modeling with evidence integration has captured perceptual and neural dynamics elicited by parametric stimuli in simple tasks, but studies of natural vision reveal richer dynamics that remain poorly understood. In this study, we propose samples in time are not accumulated from a noisy external environment, but from internal representations formed through Bayesian inference where the statistics of sensory inputs are refined iteratively. Thus, we aim to test if iterative Bayesian inference determines perceptual dynamics when processing natural stimuli. To test this, we focus on natural image segmentation. We measured human perceptual segmentation using a recently published experimental design: participants judged whether pairs of regions in an image were in the same segment (‘yes’) or not (‘no’). Subjective segmentation maps were reconstructed for each participant with optimization on ‘yes’/‘no’ responses per pair. Examining responses where perceived segments were inconsistent with the segments established by the optimal subjective map, we observed that participants presented a bias toward responding ‘yes’ when the two regions were close and ‘no’ when far. Furthermore, decision times increased with distance for ‘yes’ responses, but decreased with distance for ‘no’ responses, and this effect was larger for participants with stronger bias. For further inquiry, we developed image-computable segmentation models of the classical evidence integration and iterative Bayesian inference theories. Although both model types fit aggregate decision-time distributions similarly well, we found that the spatiotemporal dynamics observed in the data were captured only by iterative inference incorporating a Bayesian spatial proximity prior. This work highlights the importance of considering iterative Bayesian computations to understand human perceptual dynamics when exact inference is intractable, as in most real-life situations.

Topic Area: Visual Processing & Computational Vision

Extended Abstract: Full Text PDF