to s- . The precise value of the target outputs is not important, so long as they are
not too close to 0 and 1, to avoid floor or ceiling effects. Similarly, by avoiding
choosing extreme values for the units in each stimulus one has the possibility of
testing the model with stimuli that are more, or less, intense than the training ones.
The stimulus s- can be interpreted as modelling two kinds of real stimuli: an
experimental background (the so-called ‘contextual stimuli’), or a stimulus differ-
ing from s+ along a given dimension (e.g. lights of different colours or sounds
of different intensities). These two situations are usually referred to as ‘non-
differential’ and ‘differential’ training respectively, an important difference being
that the contextual stimuli can often be considered as having a very low or zero
intensity along the dimension that will be probed in the generalization test, for ex-
ample a dark key as opposed to an illuminated one when investigating wavelength
generalization, or silence in contrast to a tone when looking for sound-frequency
generalization.
From these considerations we see that the stimuli in figure 2 are best suited
to model a non differential training procedure. A differential discrimination task
can be modelled by having two stimuli such as the present s+, but differing in the
position of the signal, as we will do below.
3.2 Testing the Model
In the model, stimulus control has the same definition as in reality: the change in
response (network output) coming from a change in stimulation (changes in the
values of stimulus units). Thus, in the model, as in real experiments, a generaliza-
tion test can be given along many dimensions, and we choose here to consider (1)
the intensity of the signal in s+, that can be varied from 0 to 1 (‘intensity test’),
and (2) its position, shifting it from left to right (‘translation test’).
In the intensity test the stimulus change affects the same units throughout the
testing, while the translation test can be understood as a rearrangement of the
stimulation coming to the network while keeping constant its total intensity. A
compound case would be changing the size of the signal in s+, involving thus
both a change of total intensity and different sets of units involved by different
test stimuli. This latter case will not be investigated in this paper (but see Grice &
Saltz, 1950; Mednick & Freedman, 1960; Dougherty & Lewis, 1991, for experi-
mental data).
We have chosen the intensity and translation tests because they are readily
compared to experiments: intensity generalization gradients have been measured
along many dimensions, including sound or light intensity (Razran, 1949; Thomas
& Setzer, 1972; Zielinski & Jakubowska, 1977), colour intensity (Czaplicki et al.,
1976), taste (Tapper & Halpern, 1968), and click frequency (Weiss & Schindler,
1981); rearrangement of stimulation is implied in studies using such dimensions