To investigate the role of pre-depression rumination in the network model, it is useful to consider whether the network's performance changes as a function of whether the affective-semantic loop was allowed to function during the learning of negative stimuli. In fact, when the network was allowed to ``ruminate'' during overtraining on negative stimuli, it began to evidence negative information processing biases very quickly (8 epochs as opposed to the 100 epochs needed to achieve biases of the same magnitude in the nonruminating network). In this way, rumination may be conceived of as a vulnerability factor for depression; the same environmental stimuli affect the network much more in the presence of rumination than in its absence.
With even 20 epochs of training on negative stimuli the network begins to associate many positive stimuli with a neutral affective valence, and neutral stimuli with a negative affective valence. After 40 epochs, even positive stimuli are evaluated as negative as soon as input from the network's hidden layers is shut off, as shown in Figure 17, p. 133. The clinical analog of this phenomenon would be a person perceiving a traditionally neutral stimulus, while in a depressed mood, contemplating the stimulus, and appraising it in a negative manner. For example, if a person saw a pen, which usually has a neutral valence for them, she might quantize the valence to negativity, to return an appraisal of ``icky black thing which leaks in my pocket.'' As such, the network effectively has more difficulty at performing the valence-identification task with neutral words, with respect to negative words, often actually returning a negative word when asked for a neutral word. This added latency is reminiscent of the increased latencies on most cognitive tasks, regardless of affective content, displayed by people who are depressed.
