B7: Classification images

S Barthelmé

University of Geneva, Switzerland

A large part of vision science is about figuring out the rules that govern perceptual categorisation. What makes us see a person as male or female? A pattern as symmetric or asymmetric? A smile or a frown on a face? Classification images (Ahumada and Lovell, 1971), use noise to uncover the rules defining a perceptual category. Adding a moderate amount of noise to the picture of a smiling face will produce a random stimulus, essentially a "perturbed" version of the original: still identifiable as a face, but with altered features (Kontsevich and Tyler, 2004). Depending on the exact pattern of the noise, the perturbed face will sometimes look just as smiling as the original, sometimes distinctively less so. Viewed geometrically, what this means is that the added noise sometimes takes the original stimulus across the smiling/unsmiling boundary. The intuition behind the original technique is that by looking at those noise patterns that lead to a response change and comparing to those that do not, we should be able to characterise the features that the visual system uses to decide whether a face is smiling or not. In this tutorial I will introduce this classical technique and a number of applications, but I will focus especially on setting a broader context. Although classification images are native to psychology, they have close cousins in many areas of science (Murray, 2011). We will see that classification images have interesting ties to a range of concepts and techniques across the disciplines, from Generalised Linear Models in statistics, to compressed sensing in computer science. Putting classification images in context helps us understand why they work, when they work, and how they can be extended. Ahumada, A. J. and Lovell, J. (1971). Stimulus features in signal detection. The Journal of the Acoustical Society of America, 49(6B):1751-1756. Murray, R. F. (2011). Classification images: A review. Journal of vision, 11(5).

Up Home