Learning to See Stuff

Profile picture By: Roland W. Fleming,  Kurt Koffka Professor of Experimental Psychology, Giessen University, Germany

Thursday 13 May 2021 – 13:00 BST

Join via Zoom
Meeting ID: 915 2023 3134 | Passcode: 296251

Under typical viewing conditions, humans effortlessly recognize materials and infer their properties at a glance. Without touching materials, we can usually tell what they would feel like, and we enjoy vivid visual intuitions about how they are likely to respond if we interact with them. These achievements are impressive because the retinal image of a material results from extremely complex physical processes (e.g., sub-surface light transport; visco-elastic fluid flow). Due to their extreme diversity, mutability and complexity, materials represent a particularly challenging class of visual stimuli, so understanding how we recognize materials, estimate their properties, predict their behaviour, and interact with them could give us more general insights into visual processing. What is ‘material appearance’, and how do we measure it and model it? How are material properties estimated and represented? Discussing these questions causes us to scrutinize the basic assumptions of ‘inverse optics’ that prevail in theories of human vision, and leads us to suggest that unsupervised learning may explain aspects of how the brain infers and represents material properties.
  Consistent with this idea, I will present some recent work in which we show that an unsupervised network trained on images of surfaces spontaneously learns to disentangle reflectance, lighting and shape.  More importantly, we find that the network not only predicts the broad successes of human gloss perception, but also the specific pattern of errors that humans exhibit on an image-by-image basis.  These findings should hopefully be of interest to psychologists, neuroscientists, philosophers and AI / machine learning researchers.