Incorrect Responses

The inclusion, in the model, of ``No'' responses allowed an investigation of how well the model approximated human sensitivity to stimuli in the lexical decision task. Dprime (z(p(hit))-z(p(false alarm))) was calculated for each stimulus duration, for each type of stimulus, independently for nondepressed and depressed conditions. Whereas with people, one can not know exactly what stimuli are depressogenic for them (i.e., caused their depression), and thus which stimuli they might be expected to be very sensitive to, this information is present for the model. To evaluate the hypothesis that a depressed network would be most sensitive to depressogenic stimuli, dprime was calculated separately for depressogenic (DG) and nondepressogenic (NDG) negative stimuli. Table 13 describes the obtained dprime estimates for the simulation. Sensitivity was moderate, due to the high numbers of false alarms. As with the human data, sensitivity increased as the simulated stimulus duration increased. Like humans, for low stimulus durations, the sensitivity was better for the nondepressed network than the depressed network, except when the stimulus was depressotypic, as might be expected. That is, for the network, increased training on some stimuli had the expected benefit of allowing the network to better classify its training set. This advantage largely disappeared for a stimulus duration which was effectively infinite (i.e., all lexical determinations were made in under 500 epochs). Interestingly, this advantage disappears when the network is severely overtrained on any negative stimulus with sufficient noise, such that it attempts to classify all perceived stimuli as the depressogenic stimulus.

 

Table: Dprime for Simulated Data

Duration (epochs)	Valence	NonDepressed Dprime	Depressed Dprime
080	Negative (DG)	-2.33	-2.04
080	Negative (NDG)	-1.75	-3.03
080	Neutral	-1.54	-1.44
080	Positive	-2.52	-1.91
500	Negative (DG)	-3.31	-3.28
500	Negative (NDG)	-3.31	-3.28
500	Neutral	-3.31	-3.28
500	Positive	-3.31	-3.28