Figure 8.2 Explanatory or Causal Relationship
ARTIFICIAL NEURAL NETWORKS: THEORY AND APPLICATIONS
9.4 Theoretical foundations of ANNs
9.4.6 Learning or training method
Learning takes place when there is any change in the memory, W. Mathematically, it is defined as:
A
Learning = --- * 0
dt
All learning or training methods can be classified into two categories: supervised learning and
unsupervised learning. Supervised learning is a process that incorporates an external teacher and/or
global information. It requires pairs o f data consisting o f an input pattern and the correct result;
training data must contain the solution that the network is expected to provide, which is sometimes
difficult to obtain. A series o f case history, either actual or contrived, is therefore required. In
supervised learning, external control of the training process is also necessary, such as, deciding when
to stop the training. On the contrary , unsupervised learning, also referred to as self-organisation, is
a process that incorporates no external teacher and only relies upon local information and internal
control. Unsupervised learning self-organises presented data and discovers its emergent collective properties. It classifies input patterns internally and has no need for an expected result. Hence, the
data requirements for unsupervised learning are less demanding.
Learning methods form the backbone o f all ANN paradigms. In essence, learning models in ANN
are rules or procedures that tell a PE how to modify its synaptic weights in response to stimuli
(Caudill and Butler, 1989). The following sub-sections briefly describe some popular learning
algorithms.
9.4.6.1 Error-correction Learning
Error-correction learning is a supervised learning procedure that adjusts the connection weights
between PEs in proportion to the difference between the desired and computed values o f each PE in
the output layer. If the desired value o f the jth output layer PE is Cj and the computed value o f the
AW.. = a a . [ c ^ - ô . ]
where Wÿ is the memory connection strength from to bj; and
a is the learning rate, typically 0< a « 1.
The error-correction learning algorithm is adopted by one o f the most popular ANN paradigms,
Backpropagation. This ANN paradigm will be fully discussed in Section 9.5.3.
9.4.6.2 Reinforcement Learning
Reinforcement learning is similar to error-correction learning in that weights are reinforced for
properly performed actions and punished for poorly performed actions. The difference is that error-
correction learning requires an error value for each output layer PE to derive a vector o f error values,
and reinforcement learning requires only one value to describe the output layer's performance, that is, a scalar error value. The scalar success or failure value is provided by the environment. Bailey
and Thompson (1990) considers this learning algorithm as a compromise between supervised and
unsupervised training as it requires an input and only a grade or reward signal as an output. Hence,
training data requirements are less stringent than those for supervised learning.
9.4.6.3 Hebbian Learning
Hebbian learning, named after Donald Hebb (1949) who articulated the concept o f correlation
learning but not the mathematical formalisation, is the adjustment o f a connection weight according
to the correlation o f the values o f two PEs it connects. The original statement o f Hebb's law reads
as follows:
"When an axon o f cell A is near enough to excite a cell B and repeatedly or persistently takes p a rt in firing it, some growth process or metabolic change takes place in one or both cells such that A's efficiency, as one o f the cells firing B, is increased. "
The sim plest mathematical form of this learning rule, the simple Hebbian correlation, where the
The net efifect o f this process is that the strength o f the interconnection from A to B increases. This
implies that B will become more sensitive to A's stimulus after appropriate training has occurred.
9.4.6.4 Competitive and Cooperative Learning
Competitive and cooperative processes are described as ANNs with self-exciting recurrent
connections and neighbour-inhibiting (competitive) and/or neighbour-exciting (cooperative)
connections. They are essentially self-organising systems that can learn by themselves as the PEs
in the network compete for the privilege of learning. Competitive learning, introduced by Grossberg
(1972 and 1976), is a pattern o f classification procedures for conditioning inter-layer connections
in a two-layer ANN. In the simplest form where "winner takes all", competitive learning works in
concert with recall in the following manner;
1) an input pattern is presented to the input layer of PEs;
2) the input layer PEs send their activations through the to Fg inter-layer connections
to the Fg layer;
3) each Fg PE competes with the others by sending positive signals to itself (recurrent
self-excitation) and negative signals to all its neighbours (lateral neighbour-inhibition);
and
4) eventually the Fg PE with the greatest activation will be singularly active and all others
will be nullified. The Fg PE with the largest activation is called the Fg-winner.
Once the competition is complete, adjustment or conditioning o f the F^s^ to Fg-winner inter-layer
connections takes place. This means that only the weights o f the connections emanating from the
winning Fg PE are adjusted, leaving all other connections unaffected. In this respect, only the
winning PE is allowed to learn, which is a distinctive feature o f a competitive learning network. In