Neural computing in pharmaceutical products and process development - Computer-Aided Applications in Pharmaceutical Technology - page 93

unseen cases (data not presented during the training), it can be said that
they are used for data interpolation (Sun et al., 2003).
ANNs were fi rst introduced in the area of drug discovery in the 1980s,
specifi cally for studying the chemical problems of quantitative structure-
activity relationships (QSAR) (Zupan and Gasteiger, 1991; Livingstone
and Salt, 1995). Today, this is a widely used technique in pharmaceutical
research for prediction of nonlinear relationships between casual factors
(ANN inputs) and response variables (ANN outputs). It has been shown
that many artifi cial intelligence systems, especially neural networks, can
be applied to fundamental investigations of the effects of formulation
and process variables on a drug delivery system (Sun et al., 2003). In the
area of pharmaceutical technology, ANNs can be applied as both
classifi cation and modeling (or optimization) techniques. Of course,
some of the networks are specifi c to certain tasks.
Once the ANN model is built, during the training process, it can be
used for predictions of outputs for inputs previously unseen by the
network. It is known that careful data handling and model construction,
avoiding the known pitfalls associated with neural networks, can lead to
analyses that outperform traditional statistical methods (Manallack and
Livingstone, 1999). This is especially because the functional relation-
ship between the input and output data (independent and dependent
variables) does not need to be known prior to the network training and
it is often not revealed even after the ANN model giving satisfactory
predictions is built. A typical ANN model maps nonlinear input-output
relationships.
ANNs can be used as single as well as multi-objective simultaneous
optimization techniques that outperform response surface methodology
(Takayama et al., 2003). This means that multiple dependent variables
can be modeled simultaneously, which is a signifi cant improvement in
comparison to conventional techniques. It could be postulated that
ANNs of any number of inputs and outputs can be constructed.

￿

￿ ￿

5.1.2 Theory

ANNs are generally composed of layers of processing units - neurons
(nodes) (Figure 5.1). Artifi cial neurons are processing units that rely on
the principles of biological neuron functions. This means that the artifi cial
neuron receives information, elaborates on it, and then transmits further
to other processing units.