FEEDFORWARD NEURAL NETWORKS IN CHEMISTRY - MATHEMATICAL SYSTEMS FOR CLASSIFICATION AND PATTERN-RECOGNITION

An FFN can often, but not always, improve on existing methods of making predictions from a given data set if the size of the data set permits generalization. Unfortunately, the number of examples available for typical problems in QSAR is small: only four of the applications included more than 200 examples in the training set. In virtually every case, the number of weights exceeded the number of training examples. Smaller networks might have sufficed, but few researchers based their selection of architecture on performance. Many of these results might improve with the application of current best practice, especially cross-validation and selection of architecture based on predictive ability. Even with these problems, FFNs generally compare favorably with existing methods. © 1993, American Chemical Society. All rights reserved.