Batch or Online Learning

If the training set is not fixed because new data are being added as training is performed, or if the samples in the training set are changing with time, as is often the case when neural networks are used in control applications, online learning is essential. 

Feedforward networks of the sort described in this chapter are the type most widely used in science, but other types exist. In contrast to a feedforward 

Artificial neural networks are now widely used in science. Not only are they able to learn by inspection of data rather than having to be told what to do, but they can construct a suitable relationship between input data and the target responses without any need for a theoretical model with which to work. For example, they are able to assess absorption spectra without knowing about the underlying line shape of a spectral feature, unlike many conventional methods. 

Most recent scientific applications involve the determination of direct relationships between input parameters and a known target response. For example, Santana and co-workers have used ANNs to relate the structure of a hydrocarbon to its cetane number,4 while Berdnik s group used a theoretical model of light scattering to train a network that was then tested on flow cytometry data.5 

QSAR studies are a fertile area for ANNs and numerous papers have been published in the field. Katritzky s group has a range of interests in this area, particularly related to compounds of biological importance see for example Reference 6. Some QSAR studies have been on a heroic scale. Molnar s group has used training sets of around 13,000 compounds and a total database containing around 30,000 to try to develop meaningful links between cytotoxicity and molecular descriptors.7 

---