As for depressed humans, the depressed network is significantly delayed in responding to positive with respect to negative words on the valence identification task. Also, as in depressed humans, the depressed network is slightly delayed in responding to negative words on the lexical decision task with respect to positive or neutral words, and with respect to the nondepressed network's performance. As with the human data, the latter difference is very small (4 epochs) and would not be considered ``statistically significant'' given the relatively small sample of simulated subjects, and the relatively large variability between subjects. To support the idea that the statistical nonsignificance of this result is indeed the product of low power, a sample of 50 more simulated subjects was added to the current sample, without qualitatively changing the direction of the findings. As for humans, biases on the valence identification task are much more closely tied to induction of depression than are biases on the lexical decision task. Finally, as in the human data, the valence identification task, in general, takes longer than the lexical decision task.
| Non- | ||||||||
| Dura- | Depressed | Depressed | ||||||
| Task | tion | Valence | Mean | St Dev | N | Mean | St Dev | N |
| Lexical Decision | 80 | Positive | 138.0 | 1.58 | 50 | 137.9 | 1.81 | 50 |
| Lexical Decision | 80 | Negative (NDT) | 136.4 | 1.74 | 50 | 140.5 | 1.98 | 50 |
| Lexical Decision | 80 | Negative (DT) | 146.6 | 2.89 | 50 | 147.4 | 2.45 | 50 |
| Lexical Decision | 80 | Neutral | 143.9 | 1.89 | 50 | 145.3 | 1.99 | 50 |
| Lexical Decision | 80 | Nonword | 148.0 | 1.29 | 50 | 147.3 | 1.19 | 50 |
| Valence Identification | 80 | Positive | 125.5 | 1.76 | 50 | 165.6 | 5.52 | 50 |
| Valence Identification | 80 | Negative (NDT) | 129.0 | 4.17 | 50 | 116.8 | 2.55 | 50 |
| Valence Identification | 80 | Negative (DT) | 146.8 | 5.16 | 50 | 106.4 | 1.83 | 50 |
| Valence Identification | 80 | Neutral | 202.5 | 8.55 | 50 | 201.5 | 5.1 | 50 |
To explain the lexical decision task results it may be said that the depressotypic overtraining effectively makes the network associate stimuli close to depressotypic stimuli in representational space with these depressotypic stimuli. As all negative words share a valence with depressotypic stimuli, when the network is presented with a negative stimulus it becomes effectively ambivalent as to whether a stimulus should take the semantic representation of the stimulus it originally learned, or the semantic representation of a depressotypic stimulus. Its responses are thus particularly delayed on nondepressotypic stimuli.
To explain the valence identification task results, it may be said that overtraining the network on some negative stimuli has the effect of making these depressotypic stimuli more accessible to the network. Thus, when the network is confronted with other stimuli it tends to associate their affective valence to the affective valence of stimuli which it has learned (i.e., negativity). It thus has the effect of making positive stimuli difficult to recognize. Negative stimuli are recognized little faster than negative stimuli for nondepressed people primarily because, given the slow rate of diffusion into the network's accumulators, it is difficult for the network to ever recognize a stimulus faster than 100 epochs; the analogous explanation proposed for the lack of facilitation on negative words in depressed human performance is that humans are already performing at the lower bound of their possible performance. Valence identification takes longer than lexical decision in the network for a number of reasons. First, the process of lexical identification is done at an earlier stage than valence identification, thus incurring less noise. Second, valence identification takes activations only from lexical identifications making it difficult to perform a valence identification significantly before a lexical decision. Finally, the ``no'' threshold on the lexical decision task establishes an upper bound on the time a lexical decision will take. No such bound exists for valence identification.
Thus, at first inspection the affective interference theory, when operationalized in a neural network model, appears to mimic human performance on two information processing tasks. Interestingly, the network behaved not, as predicted by intuitive speculation regarding the task (i.e., with facilitation on negative words for the depressed condition) but as the depressed humans did (delays on positive words on the valence identification task). This finding attests to the importance of computational simulation in research. Were such simulations not to have been performed it would be easy to doubt the affective interference model, the manner in which the tasks were conducted, etc. With the added information from the network model, it seems as if humans are performing just as they would be expected to perform, and thus there is little reason to doubt the affective interference theory or the task administration. Still, as much more information that just mean reaction times was analyzed in an attempt to understand the human data, similar analyses will be performed for the network model.
