Dimensionality of Depression

To simulate different levels of depression, the network may be trained, to different degrees, on the negative stimuli. As this is the only source of variation between models, it may shed a great deal of light on variability in depression. Overtraining the model more on negative stimuli to a point had the effect of increasing the observed information processing biases (e.g., increasing observed latencies by the ``depressed'' network on negative words on the simulated lexical decision task). Yet, when the network was overtrained on negative stimuli for extremely long times, another pattern began to emerge. Specifically, the network began to make many mistakes on the tasks. All stimuli would be evaluated as negative on the valence identification task. All stimuli including nonwords would be evaluated as depressotypic words on the lexical decision task. Such a phenomenon might be likened to severe cases of clinical depression in which all a person can focus on are their depressotypic cognitions. Any incoming stimuli are related to these cognitions. For example, severely depressed people often become ``expert'' at turning any statement around such that it is interpretable as indicating them being a failure.

The increasing magnitude of information processing biases with overtraining is exemplified in Table 11, p. 112, which depicts the time the network took to identify a particular positive, neutral, nondepressotypic negative, and depressotypic stimulus on the lexical decision and valence identification task after various amounts of training. The simulated SOA for the example is 80. For the example, there is no noise in determination. As can be seen from the table, simulated reaction times to the nondepressotypic negative stimulus increase monotonically by a small amount on the lexical decision task, with an increase in depressive training. Importantly, since the increase is only one reaction-time epoch per 50 training epochs, the change is easily obscured by even a small amount of noise. Such a phenomenon might have contributed to the apparent lack of an effect of BDI on human reaction times to negative stimuli. Also as expected, reaction times to the positive stimulus increase monotonically on the valence identification task with training on the depressotypic stimuli. After 300 epochs of training, the network evaluates the positive stimulus as negative, for the first time, prompting the comparatively large increase in reaction times. This phenomenon might be likened to a person being so overcome by depression that everything looks negative to them, and they are thus unable to perform the valence identification task.

 

Table: Increasing Information Processing Biases with More Training on a Few Negative Stimuli. Values Are Reaction Times, in Epochs.

	Epochs of overtraining on depressotypic stimuli
Stimulus type	0	50	100	150	200	250	300
Lexical Decision Task
positive	116	115	115	115	115	115	115
neuteral	139	146	136	128	129	135	140
negative (NDT)	142	151	154	155	156	157	158
negative (DT)	120	115	114	114	113	113	113
Valence Identification Task
positive	124	136	144	151	157	166	392*
neutral	294	276	275	273	274	273	271
negative (NDT)	103	101	101	101	101	101	101
negative (DT)	134	106	104	103	103	102	102