More specifically, the degree to which such a model is useful is a function of what it explains, and how well it has been validated. If the model explains phenomena which have not, in the past, been understood, it is of some utility. If it can be shown that the model embodies dimensions of statistical conclusion validity (the establishment of covariation), internal validity (the establishment of causal relationships), construct validity (establishment of the constructs between which causal relationships are hypothesized), and external validity or generalizability (Cook & Campbell, 1979) then it is more likely to be of utility to its creators and others. Computational models also contribute to four types of validity discussed by Cook and Campbell (1979). Statistical conclusion validity is provided by computational simulations by allowing the exploration of mathematical formalisms for models incorporating interactions between dynamic, chaotic processes which cannot be expressed analytically. Internal validity is provided for models by showing that behaviors assumed to emerge from hypothetical processes actually do once they are rigorously specified. Construct validity is provided for a model by showing that simulated constituents of the model which contribute to a behavior, themselves behave as the processes they are meant to simulate. External validity is provided by investigating what types of behaviors are evidenced by the simulated model which were not explicitly incorporated into the model's design. If the simulation evidences behaviors for which analogs exist in humans, the model may be said to have some generality. Possibly the greatest advantage of the modeling approach to human information processing is the making explicit of causal models used by a discipline. Without attempts to specify a model, e.g., through mathematical formalisms, processes implicit in a model may be ignored, weakening the model's explanatory power. For example, Elliot's (1992) creation of the Affective Reasoner, a computer program designed to instantiate the conception of emotions presented by Ortony, Collins, and Clore (1988) revealed salient weaknesses in this conception, and its ultimate publication contained significant changes in Ortony et al.'s original theory, necessitated by the methodical specification of an implicit model.
Because the affective interference model can be phrased in terms of interacting affective and semantic structures composed of neurons in the brain, a biologically motivated model with components devoted to the representation of semantic and affective information, and feedback between these components will be described. The type of model to be used is a so-called ``neural network'' model. Neural networks and their more general cousins, the ``connectionist'' models, have gained their popularity from their predictive power (Sarle, 1994), biological congruity (Cohen & Servan-Schreiber, 1992), ability to handle noisy data (Cohen & Servan-Schreiber, 1992), and ability to model information processing tasks (McClelland, et al., 1985). Additionally, neural networks seem capable of modeling phenomena which are difficult to model using more traditional, symbolic modeling techniques, such as behaviors for which explicit governing rules are not known (Hecht-Neilson, 1990). Neural networks allow the investigation, not only, of processes currently operating in a simulated agent, but also, account for how those processes came to operate in that manner. That is, using neural networks, a model of the development of depression may be created. This methodology will therefore be a valuable tool for the investigation of the current tasks. Before presenting the actual network model, a brief synopsis of the connectionist modeling approach will be given.