0000000000246866
AUTHOR
Alice J. O'toole
Connectionist models of face processing: A survey
Abstract Connectionist models of face recognition, identification, and categorization have appeared recently in several disciplines, including psychology, computer science, and engineering. We present a review of these models with the goal of complementing a recent survey by Samal and Iyengar [Pattern Recognition25, 65–77 (1992)] of nonconnectionist approaches to the problem of the automatic face recognition. We concentrate on models that use linear autoassociative networks, nonlinear autoassociative (or compression) and/or heteroassociative backpropagation networks. One advantage of these models over some nonconnectionist approaches is that analyzable features emerge naturally from image-b…
A Widrow–Hoff Learning Rule for a Generalization of the Linear Auto-associator
Abstract A generalization of the linear auto-associator that allows for differential importance and nonindependence of both the stimuli and the units has been described previously by Abdi (1988). This model was shown to implement the general linear model of multivariate statistics. In this note, a proof is given that the Widrow–Hoff learning rule can be similarly generalized and that the weight matrix will converge to a generalized pseudo-inverse when the learning parameter is properly chosen. The value of the learning parameter is shown to be dependent only upon the (generalized) eigenvalues of the weight matrix and not upon the eigenvectors themselves. This proof provides a unified framew…
DISTATIS: The Analysis of Multiple Distance Matrices
In this paper we present a generalization of classical multidimensional scaling called DISTATIS which is a new method that can be used to compare algorithms when their outputs consist of distance matrices computed on the same set of objects. The method first evaluates the similarity between algorithms using a coefficient called the RV coefficient. From this analysis, a compromise matrix is computed which represents the best aggregate of the original matrices. In order to evaluate the differences between algorithms, the original distance matrices are then projected onto the compromise. We illustrate this method with a "toy example" in which four different "algorithms" (two computer programs …
Principal Component and Neural Network Analyses of Face Images: What Can Be Generalized in Gender Classification?
We present an overview of the major findings of the principal component analysis (pca) approach to facial analysis. In a neural network or connectionist framework, this approach is known as the linear autoassociator approach. Faces are represented as a weighted sum of macrofeatures (eigenvectors or eigenfaces) extracted from a cross-product matrix of face images. Using gender categorization as an illustration, we analyze the robustness of this type of facial representation. We show that eigenvectors representing general categorical information can be estimated using a very small set of faces and that the information they convey is generalizable to new faces of the same population and to a l…