Both peak shift and intensity generalization are natural properties of the multi-
layer perceptron. In addition, the degree of generalization is an emergent property
of the model. This is in contrast with ‘similarity’ and gradient-interaction theories
in which the degree of generalization has to be assumed.
Two very different learning algorithms, the back-propagation and artificial-
selection ones, have been used here, and the networks gave the same results irre-
spective of which one was used, with only minor quantitative differences. In ani-
mals, itis still not clear whether there are any fundamental differences between in-
nate and learned recognition (Baerends, 1982). What the network model suggests
is that very similar properties can arise from very different learning mechanisms,
given that they act on the same system (the same network in the model, the same
nervous system in reality).
Also, it should be pointed out that not all models of neural networks have the
properties of the multi-layer perceptron. For instance, it is very difficult to build
models of behaviour based on the popular Hopfield model (or derived models),
since they will not be able to reproduce any of the particular phenomena of stim-
ulus control considered here. On the other hand, some models that have not been
thought originally as artificial neural networks share some properties with multi-
layer perceptrons. This is the case, for instance, of the model in Blough (1975),
which can be considered a network without hidden layers. We feel anyway that
network models can be more readily extended to come closer to real nervous sys-
tems, in particular by more accurate modelling of neurons, synaptic dynamics and
overall organization of the network. In addition, network models force us to think
about what is the more appropriate input that has to be fed to the network in order
to accurately model the interaction between a real nervous system and physical
stimulation. This reduces the possibility of using in a model abstract represen-
tations of stimuli in terms of ill-defined ‘features’ of stimulation. For example,
Blough (1975) regarded as a drawback the fact that stimulation had to be cut in
small elements before being tractable by his model, because he couldn’t attribute
any meaning to such elements. In network models these small elements have the
very clear meaning of being the sensory cells’ reactions to stimulation (this point
is the object of a forthcoming paper).
To conclude, our results suggest that artificial neural networks can be accurate
models of behavioural phenomena, providing a valuable tool for researchers in
the field of animal behaviour. In particular, some of the concern that these models
have raised seems unjustified (see for instance Dawkins & Guilford, 1995).
12