Conclusions for Investigations of Descriptors

Consequently, statistical investigations of descriptors should be performed on vectors containing at least 128 components and training sets containing at least 50 compounds. 

Similar conditions related to the sample size apply to the investigations with neural networks, which are, in fact, nothing more than a more complex statistical algorithm. Since a network minimizes an overall error, the proportion of types of data in the set is critical. A network trained on a data set with 900 good cases and 100 bad ones will bias its decision toward good cases, as this allows the algorithm 

With an insufficient number of training data, the leave-one-out technique was applied. In this procedure all available data are used for the training of the network, except the one for which a prediction or classification has to be performed. This method can be applied iteratively for each object in the data set. 

---