Seminar – The perception of object shape in humans and machines

This talk focuses on how humans process and understand shape information.  I will present ShapeComp, a novel computational model that predicts human shape similarity judgments with high accuracy. The model, based on the statistics of natural animal shapes, outperforms traditional object recognition neural networks in capturing human-like shape representations. The lecture demonstrates the utility of ShapeComp in three key areas:

1. Automatic generation of perceptually uniform shape spaces for experimental stimuli
2. Investigation of high-level shape aftereffects in human vision
3. Potential improvements to the robustness of object recognition neural networks

The overarching goal of this work is to better understand human visual reasoning capabilities, particularly in situations with limited data, and to develop more human-like artificial intelligence systems. This research has implications for cognitive science, computer vision, and the development of more robust and generalizable AI models that can process visual information in ways more closely aligned with human perception.

E Building, ET-14.