Chapter 2. Prehension

Appendix C - Computational Neural Modelling 389
input directly from every other unit, hence the number of connections
is equal to the square of the number of units. As before, some units
are defined as input units, while others are defined as output units.
Recurrent networks are capable of more complex behavior, in
particular rather than converging on a single state, they may settle into
limit cycles of neural activity. Such networks are in general harder to
analyze, but are necessary when dynamic behavior is desired.
C.2.2 Adaptation in neural networks
Computational models can be adaptable or non-adaptable. Non-
adaptive networks have fixed synaptic weights and behavior, whereas
adaptive networks modify their synaptic weights (and therefore their
behavior) over time according to the stimuli they experience. Hybrid
nets also exist which have ‘pre-wired’ non-adaptive portions
interacting with adaptive portions, as was seen in the Kuperstein
(1988) model in Chapter 43. A non-adaptive network might model a
peripheral area which does not change its behavior over time, such as
the spinal circuitry subserving the knee-jerk reflex. Adaptive neural
nets would be used to model parts of the brain concerned with
learning, which change their behavior with experience.
In an adaptive neural network, the synaptic connection weights
modify themselves based on locally available information, that is the
‘activity levels’ in the pre- and post-synaptic neurons. This is one of
the compelling properties of neural networks - that while the individual
units are performing simple computations on the small scale, the
overall network may develop complex emergent behavior. There are
two classes of adaptation rules - supervised and unsupervised.
Supervised adaptation rules contain some measure of the desired
behavior of the network, while unsupervised adaptation rules do not.
(Thus networks governed by supervised learning develop behavior
defined by the training signal, whereas those governed by
unsupervised learning respond to properties of the patterns presented,
extracting repeated information or correlations in the data.)
In neural network models, this adaptation is captured in a learning
&. In 1949, D. 0. Hebb described a rule which has been
generalized to the following: adjust the strength of the connections
3An adaptive network can also be used as if it were non-adaptive. For example,
once the Model Network in the Jordan (1988) network learned the inverse
kinematics of the arm, it was used as a fixed network while the Sequential Network
learned the sequences.