Neural computing in pharmaceutical products and process development - Computer-Aided Applications in Pharmaceutical Technology - page 108

The most important steps in ANN construction
and testing

Figure 5.5

has already been accentuated that the number of neurons in the input
layer (i.e. the number of input variables) should be kept to its minimum.
The most diffi cult task at this stage is selection of the number of hidden
layers and the number of neurons in the hidden layer (Section 5.1.2). The
trial and error approach is still commonly used to resolve this issue, as
well as various optimization techniques.
Once the network architecture (topology) is defi ned, it is often
necessary to select the type of transfer (activation) function, learning rate,
smooth factor, etc. Some of these parameters are predefi ned in the
software used, and some need optimization. We should always bear in
mind the difference between the ANN model (architecture, topology, and
arrangement) and the ANN algorithm (method used for computation of
outputs based on inputs).
The next step is the actual training of the network. The usual practice
is to divide data into three sets: training, testing, and validation. Training
and test data sets are presented to the network fi rst, during the training
process. Training data are used to defi ne and optimize neuron weights,
whereas test data are kept aside by the network and used periodically to
check predictive ability of the network (during the training process).
Training data should cover as much as possible of the data variability. As
mentioned previously, once the test error stops decreasing or starts to
rise, it is indicative of the network overfi tting and the training process

￿

￿ ￿