Dr Yaniv Morgenstern | Towards understanding human shape representation with image-computable models

An object’s shape helps us recover what it is and how it works, even if we

have never seen it before. A major debate in cognitive science and AI

research is whether our brains achieve such abilities using ‘dumb’ but

efficient heuristics or deep but costly ‘rational’ computations, or both.

In this talk, I will present my recent work constructing ShapeComp—a

psychophysically-validated, heuristics-based model of human shape

similarity— a tool that will help us understand computations underlying

human shape perception. Merging computer vision models and Generative

Adversarial Networks from machine learning, I show how ShapeComp is highly

predictive of human shape similarity across novel shape pairs and larger

shape sets. I describe how ShapeComp can be used to synthesize carefully

controlled datasets that allow experimenters to decouple heuristic and

rational explanations, and test when deeper, more causal variables underlie

our perception of shape and its characteristics.