and the control one is due to differences in the amount of training). To model the
training of Hanson’s pigeons, the networks have been trained in two steps: first
in the situation of figure 2 (corresponding to a first conditioning stage to S+), and
then an s- was introduced consisting of a stimulus similar to the s+ in figure 2
but with a displaced location (corresponding to discrimination training between
two wavelength). Model data relative to the translation test, introduced in sec-
tion ‘Testing the Model’ above, are shown in figure 7. The overall shape of the
gradient, very different from that characteristic of intensity dimensions, is well
reproduced. Moreover, we see that the peak of the gradients is further away from
s+ the closer s- comes to it, an effect that can be considered the main charac-
teristic of Hanson’s (1959) study. We observe as well that the height of the peak
increases both in the model and the empirical data, but a thorough comparison
is prevented by the difference in training procedures (the networks are trained to
give a fixed response to s+, while this variable was not controlled in Hanson’s
(1959) experiment, as it is clear from figure 6).
5 Discussion
In this paper we have evaluated an artificial neural network model’s ability to
predict stimulus-response relationships. The model was compared with a number
of general empirical findings. In summary, the neural network model was very
accurate in predicting stimulus-response relationships, appearing, in fact, superior
to many psychological and ethological models.
Existing models of stimulus control include many theories based on similarity,
in which response is predicted to decline monotonically in all directions from, for
example, the training stimulus or a stimulus prototype. The famous bell-shaped
generalization curve, with its maximum at the training stimulus, exemplifies this.
Stimulus summation theory in ethology (Hinde, 1970) is one example of such
theory. If a stimulus contains a number of parts, removing one or several of these
is predicted to decrease the rate of responding.
Unfortunately, this kind of ‘similarity’-based theories cannot predict peak shifts
in maximum responding or supernormality. In this sense the so-called gradient
interaction theory is more refined. In addition to generalization it also models
the interactions between different experiences: for instance when an S+ (ex-
citatory stimulus) and an S- (inhibitory stimulus) are close in appearance, the
maximum responding may not occur at the S+ but for a stimulus more different
from S-. However, not all peak shifts or supernormality effects can be predicted
by gradient-interaction theory. A major phenomenon is intensity generalization,
which does not show the bell-shaped curve but a monotonically increasing rate of
response (Mackintosh, 1974).
11