3/7/24 – Cognition and Cognitive Neuroscience Brown Bag: David Coggan, PhD
Modelling human visual perception of occluded objects: insights from training 150 convolutional neural networks.
Date: Thursday, March 7
Time: 12:20 p.m. – 1:10 p.m.
Location: Wilson Hall 115
Human observers can readily perceive and recognize visual objects, even when occluding stimuli obscure much of the object from view. By contrast, state-of-the-art convolutional neural networks (CNNs) perform poorly at classifying occluded objects. In recent work, we have collected three datasets of human responses to occluded object images using behavioral, neuroimaging, and eye-tracking techniques, through which humans and computer-vision models can be compared. In this talk, I will present the key insights from evaluating a large number of CNNs that vary in their architecture, learning objectives, and visual diet. For instance, one theme that has emerged across these models is that naturalistic learning objectives (e.g. self-supervised rather than supervised) and training dataset augmentations (e.g. occlusion by real objects rather than artificial shapes) lead to more human-like patterns of performance. This suggests that human occlusion-robustness emerges in part from unsupervised engagement with the specific forms of occlusion that occur in nature.
Questions? Contact Isabel Gauthier.