COMPUTERS CAN FIND HAPPINESS*: SUPPORT VECTOR MACHINE CLASSIFICATION OF FACIAL EXPRESSIONS
Dr James M Hogan, School of Software Engineering and Data Communications, Queensland University of Technology, Australia

Abstract: The famous Ekman and Friesen "Pictures of Facial Affect" study - in which subjects were asked to categorise human facial expressions - has been of great interest to psychologists and vision researchers for many years. Of particular concern to the more computationally minded has been the identification of those features most salient in the determination of each category - both in the judgment that an expression is a prototypical example of a certain category, and in the related question of the degree of perturbation required for a categorical distinction to be made.

In recent years a number of researchers have applied neural networks of various kinds to this problem, achieving good agreement with human judgments and some useful insights into the organisation of the 'expression space'. Such studies have been aided in our case through the use of digital morphing software, allowing detailed investigation of the transitions between 'well-classified' end point images, and psychologically interesting predictions of intermediate state confusion.

Given this history - and the extraordinary success of kernel methods across a number of domains - one could be forgiven for expecting few difficulties in the application of SVMs to the problem, and even fewer observations of benefit to the wider machine intelligence community. On the contrary, image classification tasks of this subtlety serve to the highlight the role of the kernel in explicitly defining the nature of similarity between patterns, and the importance of careful selection if the constraints of the domain are to be respected. In the present case, standard kernels may provide superior classification performance, but this may often be accompanied by counter-intuitive and even absurd orderings of the non-class examples, suggesting that a purpose built kernel giving greater weight to local similarity is needed.

In this talk I shall review our neural network and support vector machine approaches to the problem, giving particular attention to the difficulties encountered with standard kernels, and the development of a novel 'receptive field' kernel based in principle upon ideas developed for string kernel methods.

(*) By the way: computers can also find Sadness, Surprise, Fear, Disgust, and Anger...

This seminar was held at the Department of Computer Science, Royal Holloway, University of London on 31 October 2002.