Previous PageNext Page

1. Introduction

In many areas where complex information needs to be processed - from the stock-market to medical diagnosis and chemistry - people are suddenly talking about "neural networks". Neural networks appear to be a new secret weapon to combat a multitude of problems. In order to illustrate this, Figure 1 shows the dramatic rise in the number of publications on the use of neural networks in chemistry. What makes neural networks so attractive? Are they really a panacea for information processing?

Fig. 1. Increase in the number of publications on the use of neural networks in chemistry over the period 1988-1991.

Neural networks were originally developed as models of information processing within the brain. This explains some of their fascination: The human brain has phenomenal processing power, far beyond that even of supercomputers today. Obviously the human brain processes information in a completely different way from that of today's conventional computers, which are constructed along the lines of the "von Neumann" architecture. A von Neumann computer works through a program (an algorithm) step by step, that is, sequentially.
In contrast the human brain operates largely in parallel: incoming information is channeled through many processing units simultaneously. This can be demonstrated by the "100 Step Paradox": We know from neurophysiology that a nerve cell or neuron recovers approximately one millisecond after firing. On the other hand we also know that the human brain is able to perform intelligent processes, such as recognizing a friend's face or reacting to some danger, in approximately one tenth of a second. Therefore the brain is able to perform difficult tasks in less than 100 sequential steps. This small number of steps is of course insufficient to solve such complex problems, so we conclude that many tasks must be performed simultaneously and in parallel.
Artificial neural networks are nowadays usually implemented as software packages which run on conventional von Neumann computers and merely simulate parallel processing. True parallel processing is only possible on appropriate hardware (Transputers) and is still rare today. The software approach permits the use of the same program for quite different knowledge domains. The same algorithm can be used to study the relationships between chemical structure and the infrared spectrum, to simulate a tennis match, or to predict stock-market trends.
In conventional programming one tries to solve problems by finding a problem-specific algorithm where every instruction is tailored to exactly the task at hand. Alternatively one may solve the problem by using an expert system, which distinguishes strictly between the knowledge specific to the task and the mechanism to draw conclusions and make decisions. Neural network algorithms, however, do not belong to any particular knowledge domain, but may be used generally to solve certain classes of problems from a wide variety of fields.
No longer is it the sequence of instructions that is specific to a task. It is the type of information that is fed into the neural network and the way it is represented there that tailor a study involving a neural network to the task at hand.
Neural networks may be used to solve the following problem types:

-

Classification: An object, characterized by various properties, is assigned to a particular category.

-

Modeling: Neural networks can output both binary and real values. Through this feature, by combining certain experimental results for an object we can arrive at other properties for it. Statistical methods produce such relationships by using an explicit mathematical equation. Neural networks however are able to express such relationships implicitly; this is especially useful in cases where an explicit equation can not be set up.

-

Association: Neural networks can be used for the task of comparing information because they are able to store information of similar kinds. Thus they are able to recognize that two pictures of a face depict the same person, even when one of the pictures is distorted (autoassociation). Furthermore, they can be used for associative tasks where one object has a particular relation to another object (heteroassociation).

-

Mapping: Complex information can be transformed into a simpler representation (e.g., projection onto a plane) while preserving all essential information.

In this introductory article we shall first introduce the basic features of the various types of neural networks before giving an overview and selected examples of the application of neural nets in the field of chemistry [1][2].

Previous PageNext Page


Johann.Gasteiger@chemie.uni-erlangen.de