Induction of Negative Information Processing Biases

While the statistical results culled from cross-sectional empirical experiments, and observations based on Bower's (1981) semantic network models presented thus far allow speculation only regarding the processes operating in a currently depressed person, the use of a neural network model allows, and in fact, necessitates that thought be devoted to how the network is to incur such information processing biases, or, in short, how the network is to become depressed. That is, if all weights in the network are assumed to begin in a relatively homogeneous configuration, and were the network above to be trained equally on stimuli with positive, negative, and neutral valences, it is expected that the network would perform similarly for positive and negative words on each of the tasks. By simulating aspects of the manner in which it is believed humans become depressed, it may be observed whether the network's performance changes in a manner congruous to the way depressed humans behave on the task.

A popular view of depression suggests that the induction of depression involves a single, pervasive negative life event or loss experience (e.g., Beck 1974, Brewin, Andrews, & Gotlib, 1993, Paykel 1979) which is continuously thought about. Such a process can be operationalized by allowing the network to overlearn one or a few negative stimuli. Specifically, depression is simulated in the network in the following manner. After it has learned all the stimuli in its lexicon to a sufficient degree (approximately 1$\%$ errors on a given presentation of all stimuli), it is trained for an additional 100 epochs on a small number of stimuli deemed to have a negative valence.

Theorists such as Lewinsohn and Hoberman (1982) suggest that induction of depression is a result of pervasive negative reinforcement in the absence of positive reinforcement. In this way, negative reactions to environmental stimuli are effectively learned. The preceding style of training would effectively simulate such an environmentally induced depression. Other theorists such as Nolen-Hoeksema (1987) suggest that much of depression is rooted in ``depressive rumination''. Such rumination could either be represented by overlearning the negative stimulus (as in the induction described above) in the absence of an actually present orthographic stimulus, or by allowing feedback between the affective and semantic representations of a stimulus to occur in the network even once the stimulus is not fed, as input to the network, during learning. Of these methods, the second appears a preferable method of representing rumination during training, as little evidence suggests that visual images of words are conjured during depressive rumination. To investigate the effects of an analog of rumination in the network, the network was allowed to repeatedly engage in the affective-semantic recurrent loop in absence of input from the stimulus, during the ``depression induction'' stage.