Encyclopedia of Computer Science and Technology

nonprocedural languages 337tions), each of which are presented with weighted inputs(called vectors) from which it calculates an output valueof either true (1) or false (0). The designer of the systemknows what the correct output should be. If a given element(or node) produces the correct output, no changes aremade. If it produces the wrong output, however, the weightsgiven for each input are changed by some increment, plusa further adjustment or “bias” factor. This adjustment isrepeated for all units as necessary until the output is correct.In other words, each neuron is constantly adaptingthe way it evaluates its inputs and thus its output, and thatoutput is in turn being fed into the evaluation process ofthe neighboring neurons. (In practice, a neural network canhave several layers of processing units, with one layer providinginputs to the next.)For example, suppose a neural network is being trainedto recognize objects based on the light being received froman array of sensors. The sensor readings are interpreted bya number of “neurons,” which should output 1 if part of thedesired object exists at the location scanned by its sensor.At first there will be many false readings—points at whichpart of the object is not recognized, or is falsely recognized.However, after many cycles of adjustment this “supervisedlearning” process results in a neural network that has ahigh probability of being able to identify all objects of agiven general form. What is significant here is that a generalizedability has been achieved, and it has emerged withoutany specific programming being required!In a computer neural network the “neurons” or nodes are “trained”to detect a pattern by being reinforced when they successfully registerit.Neural networks have been making their way into commercialapplications. They can be used to help robots recognizethe key components of their environment (see roboticsand computer vision), for interpreting spoken language(see speech recognition and synthesis), and for problemsin classification and statistical analysis (see data mining).In general, the neural network approach is most useful forapplications where there is no clear algorithmic approachpossible—in other words, applications that deal with theoften “fuzzy” realities of daily life (see fuzzy logic).Further ReadingBishop, Christopher M. Neural Networks for Pattern Recognition.New York: Oxford University Press, 1995.Haykin, Simon. Neural Networks: A Comprehensive Foundation. 2nded. Upper Saddle River, N.J.: Prentice Hall, 1998.McNellis, Paul D. Neural Networks in Finance: Gaining PredictiveEdge in the Market. Burlington, Mass.: Elsevier AcademicPress, 2005.“Neural Networks & Connectionist Systems.” Association for theAdvancement of Artificial Intelligence. Available online. URL:http://www.aaai.org/AITopics/html/neural.html. AccessedAugust 16, 2007.nonprocedural languagesMost computer languages are designed to facilitate theprogrammer declaring suitable variables and other datastructures, then encoding one or more procedures formanipulating the data to achieve the desired result (seedata types and procedures and functions). A furtherrefinement is to join data and data manipulation proceduresinto objects (see object-oriented programming).However, since the earliest days of computing, programmersand language designers have tried to create higherlevel,more abstract ways to specify what a program shoulddo. Such higher-level specifications are, after all, easier forpeople to understand. And if the computer can do the jobof translating a high-level specification such as “Find allthe customers who haven’t bought anything in 30 days andsend them this e-mail message” into the appropriate proceduralsteps, people will be able to spend less time codingand debugging the program.It is actually best to think of a continuum that has atone end highly detailed procedures (see assembler) and atthe other end an English-like syntax like that given above.Already in an early language like FORTRAN the emphasisis moving away from the details of how you multiply numbersand store the result to simply specifying the operationmuch like the way a mathematician would write it on ablackboard. such as T = I + M. COBOL can render suchspecifications even more readable, albeit verbose: ADD I TOM GIVING T, for example. However, these languages arestill essentially procedural.Some languages are less procedural in that they hidemost of the details (or subprocedures) involved in carryingout the desired operation. For example, in modern databaselanguages such as SQL what would be a procedure (or a setof procedures) in some languages is treated as a query at ahigh level (see SQL). For example: