Cog Lunch: Yudi Xie

March 17, 2026
12pm
Location: 46-3189

Zoom: https://mit.zoom.us/j/92872622193

Speaker: Yudi Xie

Affiliation: Dicarlo Lab

Title: Can human visual occlusion reasoning be explained by purely feedforward mechanisms?

Abstract: Humans have a remarkable ability to recognize partially occluded objects. Traditionally, this ability has been believed to involve world knowledge of object shapes and how objects occlude one another. However, what kind of computational models captures how the brain implement this process remains unclear. We introduce a task probing human visual reasoning of occluded objects based on global shapes. Solving this task intuitively involves strategies such as imagining compatible shapes and ruling out alternative hypotheses. We found that purely feedforward convolutional neural networks (CNNs) can solve this task and generalize to novel conditions at human-level accuracy. Furthermore, CNNs exhibit human inductive biases and are more human-aligned than the ideal observer model, despite not explicitly trained to do so. Our findings challenge the prevailing views that recurrent processing or explicit generative models are needed, and showed that CNNs and purely feedforward mechanisms can be powerful candidate models of human occlusion reasoning.