Chemometric Methods for Biomedical Raman Spectroscopy and Imaging - Emerging Raman Applications and Techniques in Biomedical and Pharmaceutical Fields - page 190

Fig. 8.5. Schematic diagram of neural network architecture. The hidden and output
layer nodes perform nonlinear operations on their inputs. Some of the common
nonlinear operations include computing hyperbolic tangents, step functions, and
sigmoid functions. During training, features corresponding to different objects from
the training set are presented as inputs and the (known) object class for each training
object is forced to be the result of the output layer. The lines connecting different
layers are associated weights. The process of learning (training) involves computing
these weights using a set of fixed values at the input and output layers. During
testing, features corresponding to an unknown object are presented at the input
and the weights from training are used to propagate the inputs through the hidden
layer all the way to the output. The output layer finally settles at a set of values
corresponding to a final classification result
which spectral features are being used for classification and which are worth-
less. The classifier operates much like a “black box”; hence, their validation
must be exceedingly rigorous. Another common problem is that of over- and
under-training. Over-training, in which the training set is predicted well but
the validation set is not because of modeling peculiarities or noise in the
training set, is easily possible with the high dimensionality of data commonly
encountered and appropriate dimensionality reduction much be implemented
[99]. Hence, training results may be excellent but validation often falls short.
On the other hand, if the network is not trained enough it is possible that
the classification may not achieve its true potential. Unfortunately, there is
no general optimal framework for choosing the amount of training required.
The algorithms implementing ANNs usually involve an error-correction tech-
nique [100] to optimize the network. Consequently, there is a possibility of the
optimization reaching only a local optimal point (minima) and it may not be
easy to know if the solution is optimal. A practical disadvantage is that ANNs
may be slow if they are too complicated, especially in the training phase but
may also be slow in the application phase.