Previous PageNext Page

11. Summary and Outlook

Although the development of computational models of information processing in the human brain can look back on a history of almost 50 years, it is only in the last four to five years that neural networks have been applied widely to problem solving. New, efficient algorithms have paved the way for a multitude of applications, and the use of neural networks is still increasing - also in the field of chemistry - by leaps and bounds.
The models of neural networks and their potential applications of classification, modeling, association, and mapping are as diverse as the capabilities of the human brain are varied. The potential in neural networks for the processing of chemical information is very far from being exhausted.
In chemistry the task is often to assign objects to certain categories or to predict the characteristics of objects. This accounts for the dominance of the back-propagation algorithm. A whole series of other neural network models exists, however, which could be applied successfully to the field of chemistry. This should be explored more widely in future.
Undoubtedly, many of the problems that have been approached with neural networks could also have been solved with statistical or pattern-recognition methods. Neural networks, however, offer capacities which exceed the possibilities of traditional methods of data analysis. Of special importance is the fact that the relationship between input and output data need not be specified in mathematical form, but is derived from the data themselves and represented implicitly. This enables the modeling even of nonlinear relationships.
The use of neural networks still requires much experimentation; guidelines to arrive as quickly as possible at a viable solution to a problem become apparent only gradually. Of crucial importance to the successful application of a neural network is the strategy for data representation; the better the chemical information to be examined is coded, the easier and better the problem may be solved.

The cooperation between our two working teams depended crucially on support from the Bundesministerium für Forschung und Technologie of the Federal Republic of Germany and from the Slovenian Research Ministry. Particularly the provision of a guest-professorship for Jure Zupan at the Technical University of Munich within the framework of BMFT project 08G 3226 has enabled us to develop a productive scientific exchange. We would like to thank especially our co-workers A. Fröhlich, P. Hofmann, X. Li, J. Sadowski, K.-P. Schulz, V. Simon, and M. Novic, who undertook with us the first steps into the Terra Nova of applying neural networks to chemical problems.

Received: July 20, 1992 [A 899 IE]
German version: Angew. Chem. 1993, 105, 510
Translated by Dr. P. Brittain, Garching (FRG)

Previous PageNext Page


Johann.Gasteiger@chemie.uni-erlangen.de