Artificial neural networks as models of stimulus control*

Figure 6: Gradients obtained in a light-wavelength generalization test with pigeons, repro-
duced from Hanson (1959). The lower gradient is a control one, obtained after training
on a 560-nm wavelength. The other gradients result from discrimination training between
the 560-nm light and other colours, as indicated in the legend.

Stimulus location

Figure 7: Network generalization gradients coming from the translation test. The s⁺ was
fixed, and four different negative stimuli were used, consisitng in a ‘signal’ like that in
s⁺ (see figure 2), but with a displaced position (indicated in the legend above). The evo-
lutionary algorithm gave entirely similar results, but for clarity only data from networks
trained with the back-propagation algorithm are shown. The input layer of the networks
was made of fifty units, and the positive stimulus was centered at position 30 (with a tar-
get output of 0.7) for the no-s^- group and all discrimination groups. The negative stimuli
(target output 0.2) were centered at the positions indicated in the legend.

10