We show that Gabor filter representations give quantitatively indistinguishable results for classification of facial expressions as local PCA representations, in contrast to other recent work. We also show that a simple discriminant analysis automatically locates regions roughly corresponding to relevant Facial Actions. Finally, we in troduce a method that typically boosts generalization performance 9% by "peeking" at all of the unlabeled training patt erns before classifying them.

Pre-2018 CSE ID: CS1999-0629

A previous experiment tested subjects' new/old judgments of previously-studied faces, distractors, and morphs between pairs of studied parents. We examine the extent to which models based on principal component analysis (eigenfaces) can predict human recognition of studied faces and false alarms to the distractors and morphs. We also compare eigenface models to the predictions of previous models based on the positions of faces in a multidimensional "face space" derived from a multidimensional scaling (MDS) of human similarity ratings. We find that the error in reconstructing a test face from its position in an "eigenface space" provides a good overall prediction of human familiarity ratings. However, the model has difficulty accounting for the fact that humans false alarm to morphs with similar parents more frequently than they false alarm to morphs with dissimilar parents. We ascribe this to the limitations of the simple reconstruction error-based model. We then outline preliminary work to improve the fine-grained fit within the eigenface-based modeling framework, and discuss the results' implicadons for exemplar- and face space-based models of face processing.

A considerable body of evidence from prosopagnosia, a deficit in face recognition dissociable from nonface object recognition, indicates that the visual system devotes a specialized functional area to mechanisms appropriate for face processing. We present a modular neural network composed of two "expert" networks and one mediating "gate" network with the task of learning to recognize the faces of 12 individuals and classifying 36 nonface objects as members of one of three classes. While learning the task, the network tends to divide labor between the two expert modules, with one expert specializing in face processing and the other specializing in nonface object processing. After training, we observe the network's performance on a test set as one of the experts is progressively damaged. The results roughly agree with data reported for prosopagnosic patients: as damage to the "face" expert increases, the network's face recognition performance decreases dramatically while its object classification performance drops slowly. We conclude that data-driven competitive learning between two unbiased functional units can give rise to localized face processing, and that selective damage in such a system could underlie prosopagnosia.