These features give rise to a hypothesis testing methodology, to be used in the model presented here, which involves the following steps. First a model of normal functioning is created without respect to a pathology. Next, a hypothesis is formed regarding the neurological, cognitive, or behavioral basis for a psychopathology, possibly with little regard for the existing model of normal functioning. This hypothesis may then be ``tested'' by perturbing the model of normal functioning in a manner representative of the hypothesis. In the current thesis, the model of normal functioning is represented by a model of intact semantic and valence recognition processes. Theories of the onset of depression will be presented, and the network will be accordingly perturbed in a manner analogous to these methods. The network`s performance will be tested by examining how the pathologized network functions with respect to how the nonpathologized network functioned. Various measures of dysfunction or abnormality in the network's processing may be examined, such as the resulting accuracy of the network's associations or the number of processing cycles or ``epochs'' the network needs to accomplish some associative task for which it has been trained. If the network's performance with respect to its nonpathological performance is analogous to the performance of a human with a given psychopathology, with respect to the performance of a human without that psychopathology, a number of conclusions may be drawn. Extrapolation or external validity (Cook, 1990) is provided for the original nonpathologized model, since, it has demonstrated human-like characteristics for which it was not designed. Similarly, validity in the form of proximal similarity (Cook, 1990) is provided for the theory of pathology since, without specifically designing a network to perform in a manner displayed by humans with a given pathology, the perturbation has resulted in that functioning. Finally, other behaviors of the network or perturbations may be investigated so as to provide new predictions about human behaviors.
The utility of this type of model may still be considered to be dubious, as neo-Popperian ideology dictates that a theory's value is dependent on its falsifiability. In many cases, it is not clear how a neural network model may aid in the falsification of a theory, given that failure of a model to mimic human performance could either reflect an inaccurate computational operationalization of the researcher's assumptions or an actual incongruity in the causal theory. Perhaps though, restricting the usefulness of neural network models to the falsification of a causal theory is not ultimately productive. Papers describing neural network models of psychopathology present a different picture of their usefulness. The models are often used as spring-boards for theorizing about the theoretical mechanisms leading to psychopathology in a manner akin to a partitioning of variance model. For example, if a neural network model of normal functioning accounts poorly for behavior by people with a pathology, and a model which incorporates a mechanism assumed to represent a factor present in people with the pathology appears to better account for the behavior in the desired population, it may be said that the altered model is a more representative model of the causal processes involved in the behavior than the original model, for the population in question. Similarly, if a researcher creates multiple models, representing multiple hypotheses regarding the origin of a pathology, the researcher may judge between the models' performance, adding assumptions to models as necessary, to suggest which causal assumptions best account for the observed behavior, the researcher is performing an operation analogous to the incremental addition of features at various stages of a hierarchical regression.