ABSTRACT
In this chapter we propose a generalized version of the classical linear autoassociator that can he shown to implement a generalized least-squares approximation under linear constraints. The standard linear autoassociator is known to implement principal component analysis, whereas the generalized model implements the general linear model (e.g., canonical correlation). In practical terms, this generalization allows for the imposition of a priori constraints that enable differential weighting of both individual units of the input code and individual stimuli. As an illustration of the utility of the generalized model, we present simulations comparing the accuracy and learning speed of the standard and generalized versions of the autoassociator for the problem of categorizing faces by sex. We show that while the two models are equally accurate, the generalized model learns the task considerably faster than does the standard model.
