Abstract - Neural Nets and Scientific Discovery: A Match Made in AI Heaven

     Ilana Marks discusses some of the current uses of neural network technology in the process of scientific research and development. A key to successful research is to be able to recognize patterns in the data and to be able to make accurate and useful predictions. However, at the same time, producing results with sufficient speed is desired. Neural networks provide all of these benefits. Several cases of neural network use in scientific projects are detailed including cancer prognosis prediction, the detection of dangerous chemical agents, immunology, and speech aids.

Introduction
     The human brain is an invaluable and amazing organ. Electrical and chemical messages bouncing around in the brain lead to actions, speech, thoughts, and all other physical and mental properties that define the human condition. One of the most interesting abilities that the brain affords us is the ability to recognize things, people, scents, and other physical stimuli. When you think about the diversity even within a certain classification of object, it becomes more amazing. For instance, within the classification of tree, there is a wide variety of different types of trees, many of which look drastically different. A pine tree looks very different from an oak tree and yet we realize that they are both trees. We recognize general characteristics of a tree and then are able to expand our definition when other specimens are encountered. We also recognize different states in the life of a tree - for example, we recognize that leaves may change color or fall off. Despite such seasonal anomalies, a tree is clearly recognizable.
     In addition to the ability to recognize things, the brain also allows us to predict outcomes and occurrences. When dark clouds appear in the sky, we can predict that rain is imminent (unless, of course, the weather report says it will rain - in that case, the sky will suddenly become clear!) Sports fans try to predict the outcome of a game based on the previous performance of the teams. Given a set of circumstances, it is possible to deduce the most likely result. Typically, the larger the set of circumstances available to make a prediction on, the more likely it is to come up with the correct results.
     Recognition and prediction are both highly involved in scientific research. Since science is based on other science it is important to recognize where a result is reminiscent of a previous discovery. Recognizing those connections allows the researcher to expand their understanding by applying information garnered from previous research. A cornerstone of good scientific research is the formulation of an appropriate hypothesis. A hypothesis is nothing more than a prediction

about what will occur. Therefore, prediction is a backbone of research. Scientists are constantly striving to increase their knowledge base in order to make more accurate predictions about new experiments.

Neural Nets and Research
     Recognition and prediction may be important to good research, however in this day and age, speed is almost more important. Universities demand that professors publish papers often and the public demands new drugs and technology. The time needed for scientists to learn all about a variety of different fields makes it prohibitive for them to make truly accurate predictions. In addition, a certain type of scientist may not recognize a certain occurrence because it may fit more in the knowledge base of another type of scientist. With the vast increase in scientific discovery over the last several decades, a mountain of knowledge is quickly being created