Title | ||
---|---|---|
A novel algorithm for local minimum escape in back-propagation neural networks: application to the interpretation of matrix isolation infrared spectra |
Abstract | ||
---|---|---|
Back-propagation training of feed-forward neural networks often results in convergence to local minima, especially when multioutput networks and large training sets are employed. These local minima manifest themselves in a small number of residual output errors >95% (opposite errors). A procedure called "Flashcard Algorithm" has been developed in this research to overcome the opposite errors through overrepresentation of the difficult examples. A new and more objective criterion also defines a sufficiently low output error to avoid unnecessary training extensions. The problem of too slow convergence or untrainable network conditions has been addressed as well. It is now possible to predict a "doomed" run within the first few epochs in order to correct the number of hidden nodes and other network parameters. Applying the flashcard algorithm to train a neural network with matrix isolation infrared spectra resulted in 100% successful prediction of bond types and functional groups from training sets of 194 and 609 spectra, respectively. In addition, training times were cut drastically and reproducibility improved by more than an order of magnitude. |
Year | DOI | Venue |
---|---|---|
1994 | 10.1021/ci00020a037 | JOURNAL OF CHEMICAL INFORMATION AND COMPUTER SCIENCES |
Keywords | Field | DocType |
infrared spectra | Infrared spectroscopy,Matrix isolation,Algorithm,Back propagation neural network,Mathematics | Journal |
Volume | Issue | ISSN |
34 | 4 | 0095-2338 |
Citations | PageRank | References |
5 | 0.94 | 5 |
Authors | ||
2 |
Name | Order | Citations | PageRank |
---|---|---|---|
Christoph Klawun | 1 | 18 | 2.63 |
Charles L. Wilkins | 2 | 58 | 10.49 |