On the error criteria in neural networks as a tool for human classification modelling

Multi layer perceptrons (MLPs) can be applied as a tool to model human classification behaviour. In the present theoretical study we attempt to interpret MLPs within the framework of mathematical psychological models for human classification behaviour, more specifically the general recognition theory and the generalized context model. Next, four error criteria are discussed that can be used in training and test of the MLPs, in relation to two types of data representation: in terms of individual deterministic responses or in terms of probabilistic responses. All error measures considered are additive, i.e. can be written as a sum across individual stimuli. It is shown that some of these error measures have very different properties given a training set, and that the interpretation of the MLP as a means to provide knowledge about the underlying human decision process depends on the complexity of the MLP-topology