aam9n@uvaee.ee.virginia.EDU (Ali Minai) (04/03/89)
At the risk of appearing naive, let me throw out a very basic question: FOR WHAT ARE NEURAL NETWORKS USEFUL? I realize that the question has many answers, as borne out by the volume of research in the area. My interest, however, is in a more rigorous characterization. It is no secret that many so called "applications" reported for neural nets are really examples of problems fitted to a solution which is, at times, much "worse" than other perfectly simple solutions that already exist. After some cursory thought on the subject, it seems to me that there can be only two reasons for using neural nets to solve a problem: 1) They offer a computational, analytical or pedagogical advantage. 2) They enable us to explore, model or understand real neural processes. In case (2), of course, efficiency etc. are of little consequence. My question applies only to case (1), so let me rephrase it: ----------------------------------------------------------------------- | What are the computational, analytical or padagogical advantages of | | neural networks AS A TOOL for modelling and problem solving? | ----------------------------------------------------------------------- As a tentative list, let me throw out one example of a possible advantage of neural networks AS A CLASS: They offer a receptive and versatile GENERAL MODEL which can be molded to fit many problems. This is not necessarily an unmixed blessing, of course, but, first of all, is this a valid statement? And if so, how far does this validity stretch? For example, can we use neural nets as tools for a first-cut induction? Theory-formation? Why not? What advantages and disadvantages do neural nets have over already extant tools such as regression, polynomial approximation etc.? This is just one example of the kinds of questions I have in mind. I appreciate that the term "neural nets" is too general, but as long as we clearly state what kind of network we are talking about, I think the question applies to all models. Other strengths that neural nets are supposed to show, such as inherent fault tolerance, distributivity of representation, adaptivity etc. should, I think, be measured against the disadvantages which stem more or less from the same sources as the pluses. The system is fault-tolerant because it is redundant, but the redundancy makes generalization more difficult. It is adaptive because it has many degrees of freedom, but this is also what makes it less efficient than a customized system. So what are the clear advantages, and what are the trade-offs? Any comments. Ali Minai Dept. of Electrical Engg. Thornton Hall University of Virginia Charlottesville, VA 22901 aam9n@uvaee.ee.Virginia.EDU
johne@astroatc.UUCP (Jonathan Eckrich) (04/08/89)
(Ali Minai) writes: >FOR WHAT ARE NEURAL NETWORKS USEFUL? I know that you are looking for more specifics to this question than I am able to supply. I can, however, tell you how in my experience, I find NNs useful. When I was in school, I attended a biomedical class where we studied pattern recognition techniques. One of the problems that the instructor made sure we would have exposure to was that of finding the 'right' paramters to measure. No matter what the pattern is, with conventional methods of pattern recognition, the prerequisite work to each specific problem is to discover and catagorize parameters that are necessary to suitably recognize all possible patterns in the defined pattern space. For example, with the task of recognizing human faces, certain parameters or characteristics of faces in general must be established. Depending on the variety of faces that are to be used (age, sex, ethnic origin), this may not be a trivial task. With a NN approach, this data only need be presented to the net, and it will in essence 'see' and use those parameters that lead toward discrimination of different patterns, while ignoring the others. Of course, the approriate net structure will have to be used. In addition, the 'best' activation and weight modification algorithms will need to be developed. I don't know, but maybe this just puts the problem into a dif- ferent form. Someone more knoledgeable than I would have to answer that. ------ Ignore any spelling errors. I quickly proof read this. -- (rutgers, ames)!uwvax!astroatc!johne nicmad!astroatc!johne