Title: Finding Scalable Principles of Vision: Computational Mechanisms of 3D Surface Slant Estimation
Speaker: Dr.Can Oluk
Date: 2 April, Thursday 16:30
This is an online seminar. To request event details please send a message to department.
Abstract: How does the human brain transform chaotic visual input into a stable, coherent understanding of the world? My research investigates the computational mechanisms underlying this transformation. Specifically, I aim to identify principles that generalize to real-world conditions, overcoming the limitations of models developed in simplified laboratory settings. To do this, I combine normative, interpretable, image-computable modeling with behavioral experiments featuring realistic complexity.
In this talk, I illustrate this approach with a study on 3D surface slant estimation. To capture real-world conditions, we increase stimulus complexity by presenting densely textured planes and using ray tracing to render precise geometric distortions. While human performance is consistent with simple models assuming locally flat surfaces, it also matches an ideal observer model using complex pattern disparities. Given the much greater precision of this optimal approach, these results suggest that the visual system likely evolved to exploit these complex cues.
I conclude by drawing examples from my other projects to highlight the broader promise of this approach, and outlining future plans to extend it to visual target identification.
Bio: Can Oluk is a postdoctoral researcher at the Brain Mind Institute at EPFL, working with Prof. Michael Herzog. Previously, he was a postdoctoral fellow at ENS-PSL with Prof. Pascal Mamassian. He earned his Ph.D. in Psychology at the University of Texas at Austin, where he completed the interdisciplinary training program at the Center for Perceptual Systems under Prof. Wilson Geisler. He holds a B.A. in Psychology from Bilkent University. His research investigates the computational mechanisms underlying human vision. He develops normative, interpretable models, tests them against behavioral data, and links these computations directly to neural mechanisms. Ultimately, he aims to identify scalable principles of visual processing that generalize to real-world complexity.