ÉÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍ» º º º NEURAL NETWORK SOFTWARE º º Give your PC a mind of its own º º º º º ÈÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍͼ Its late at night in the Computer Lab. The tall thin programmer known only as SuperTech makes a few final changes to his source code, and compiles it for the hundredth time that night. Now there are no bugs. He executes the program and turns on the Voice Synthesizer. Strange sounds begin to come from the speaker: BAAA! GOO! GAAA! DAA! DAA! DAAA...DY! It sounds like a newborn baby, trying to learn to talk! SuperTech exits the lab, leaving his program running all night. He slips into his red Porsche and drives somewhere to a luxurious super-modern house in the Valley. He sleeps the deep sleep of the satisfied inventor. He is sure of what to expect in the morning. The next day, when he enters the lab, a dozen scientists are clustered around his computer. The computer is speaking perfectly. While SuperTech slept, his computer - aided by his Neural Network - taught itself to read and speak perfectly. There are congratulations all around. SuperTech smiles at his co-workers. It is a technological triumph! Science Fiction? No, this has already happened! The computer scientists who performed this astounding technological breakthrough were Terrence Sejnowsk and Charles Rosenberg working at the Princeton University AI Lab. Their program, NET-TALK, beginning with no knowledge whatsoever, learns good pronunciation and speech control automatically, in a single overnight training session. A Neural Network is a biological model of a human brain, simulated in the binary memory of your PC. It is made up of artificial Neurons, connected to each other by Axions. Each Neuron can have many inputs, but only one output. As a Neuron gets energized by input, it fires, sending energy along axions to other nearbye Neurons. If another Neuron receives energy from two or more axions, it also will fire, propagating the excitation to others. There are hundreds or even thousands of such Neurons, arranged in layers, and all together they form a Neural Network, capable of learning from experience. The technology has advanced so rapidly that there are now over 100 companies offering Neural Network products, ranging from AT&T to startups with names like NeuralWare, NeuralTech and Neurogen. Numerous systems are already being beta-tested commercially, including systems to diagnose diseases, to determine credit ratings, to analyse radar signals, and even to compose music. American Express will be using Optical Scanners with a hard-wired Neural Network (trained to read handwriting) to read millions of Credit Card Charge Slips each day. The Department of Defense plans to spend 400 million over the next eight years to develop neural networks for defense. Three or four contracts worth several million dollars each to expand neural-net research will soon be awarded by the Air Force. The goal is to create an aircraft that can "learn" or "adapt" to its environment - for example, reconfiguring itself if controls are damaged. It will be know as a "Self-Repairing Flight Control Program". As might be expected, a major competition is underway between the U.S. and Japan to see who will triumph in this new high-tech arena. "The same kind of excitement that surrounded artificial intelligence some years ago seems to be around Neural Networks today, " says Bell Laboratories president of research. "Neural Networks are moving faster from concepts to serious applications than artificial intelligence did." About a dozen small startups have sold over ten thousand Neural Network simulation programs during the past year, allowing researchers to develop prototype applications rapidly on digital computers such as the IBM/PC. Five or six larger Corporations are offering NeuroComputers, in which the hardware differs radically from a standard PC, and is hard-wired to resemble the brain. One industry Guru believes that NeuroComputers may lead to a multibillion-dollar market for new types of chips and computers. Typical examples are the Anza Plus, a $15,000 NeuroComputer from Hecht-Nielsen Corp here in the U.S., and the PC9800, a PC Compatible NeuroComputer for $11,000 from NEC in Japan. Commercial Neural Network Systems can run up to sixty thousand dollars, but a number of companies are also offering simpler systems that you can use to experiment with Neural Networks on your own PC. NEURON EXPERT ============= NEURON EXPERT (Thinking Software, Inc. 46-16 65th Place, Woodside, New York 11377 $59.95, with extensive documentation, free AI Demo Disk, and the AI Catalog featuring over 50 other AI Programs.) Neuron Expert is a true neural network environment simulated in the standard digital hardware of your PC. It is a combined Expert System and Neural Network, and no knowlege of programming is required. It is unique among Expert Systems in that there is no need to encode the expert's knowledge into a complex series of IF/THEN rules. Instead, you enter the questions you want the network to ask the user, and all possible solutions the network can select from at the conclusion of the consultation. Then, you TRAIN THE NETWORK, by answering all the questions, and choosing the correct solution. Only ONE training session is required for each solution - the network learns quickly. When you type NEURON at the DOS prompt, you are presented with this Menu: ÉÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍ» º NEURON EXPERT º º The Neural Network Expert System º º º º 1. Demo Consultation º º 2. Quick Reference º º 3. Enter Queries º º 4. Enter Solutions º º 5. Train The Network º º 6. Expert Consultation º º 7. Tutorial º º 8. Exit º º º º Please Enter Your Choice: º º º º (c) 1988 Thinking Software, Inc. º ÈÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍÍͼ Let's go through the various Menu Choices: 1. DEMO CONSULTATION. Immediately acquaints you with a ready-to-run application. The Neuron Expert asks you questions about your boss, and at the conclusion of the consultation will tell you if he is the greatest boss alive, an ordinary boss with some good and some bad qualities, or an out and out rotten slavedriver! 2. QUICK REFERENCE. A handy online summary of all you need to know to build your own Neural Network. 3. ENTER QUERIES. Enter up to 25 Queries, terminating each with a Question Mark, and then press enter. After your last query, press enter to return to the main menu. Use query zero(0) as a description of your system (as in the demo). You have 256 characters for each query. 4. ENTER SOLUTIONS. Enter up to three solutions, terminating each with a period. If you only have two solutions (like YES and NO) just press enter when asked for the third. You have 256 bytes to use for each. 5. TRAIN THE NETWORK. Your first solution is presented in a window at the top of the screen. Answer your Queries in a way that would lead to the first solution being true. Your answers are stored in the Network. Then your second and third solutions are processed in the same way, and the trained Neural Network is automatically stored. 6. EXPERT CONSULTATION. Now the User answers your questions, and even if his answers don't match yours exactly, the best solution is given. Neuron Expert is able to generalize to the best acceptable solution. It is rarely - if ever - incorrect. 7. TUTORIAL. Detailed information on the history and current state of the art in Neural Networks. NEURON EXPERT PROFESSIONAL II ============================= NEURON EXPERT PROFESSIONAL II (Thinking Software, Inc. 46-16 65th Place, Woodside, New York 11377 $99.95, with extensive documentation, linkable object modult, free AI Demo Disk, and the AI Catalog featuring over 50 other AI Programs.) Neuron Expert Professional II is the professional version of Neuron Expert. It increases the maximum number of queries to 100, and the maximum number of solutions to ten - enough for many complex real-world applications. It can also be included as part of your own computer program, written in any Microsoft compiler language which produces standard linkable object modules. Included on Program Disk 1 is CONSULT.OBJ, a standard linkable object module in Microsoft format, which can be seemlessly integrated into your application program with a simple CALL CONSULT (param1,param2). Param1 is the name of your trained Neural Network, say CREDIT, or CAR, or DOCTOR. (This file is automatically created during the training process). Param2 is the Solution, passed back to your program after the Consultation for possible further processing by your application. Param1 should be defined as an eight character string, and Param2 as a 256 character string. It is now possible to include a credit-risk analysis within your financial application, for example. The possibilities are endless. NEUROSHELL ========== NEUROSHELL (Ward Systems Group, available from Thinking Software, 46-16 65th Place, Woodside, N.Y 11377. $195 with paperback manual and demo systems for Credit Risk, Color Combinations, Presidents of U.S., etc) NeuroShell is advertised as "Taking Neural Networks out of the lab and into the Office", but any business problem must be chosen carefully and tested completely prior to implementation. To quote from the NeuroShell Manual: "The Neural Network technology utilized in NeuroShell is by nature a "fuzzy logic" system which is not designed to provide precise answers and solutions to every problem. It is a tool which attempts to classify patterns according to other patterns it has learned and to give the most reasonable answer based upon the variety of learned patterns. It is not guaranteed to always give an absolutely "correct" answer, especially if patterns are in some way incomplete or conflicting. Results should be evaluated in terms of the percentage of answers that match those an expert might give. In this regard, the technology is similar to biological neural functioning, after which it was designed, and differs significantly from all other conventional computer software. NeuroShell may not work at all with some applications. Some problems are well suited for the "fuzzy" logic capabilities of a neural network and others are best solved with more traditional methods." Although there are other commercially available systems that allow you to build a neural network, Ward Systems Group created NeuroShell because they believed that most of the other systems were too complex to use easily and apply to routine problems. NeuroShell has no need of computer programming expertise. It allows the knowledge engineer to concentrate on the problem and how the network technology is applied to solve it, rather than on the technical details of the Neural Network Program itself. Neural Networks have been utilized in finding patterns in signals, performing image processing and recognition, robotics, optical character input analysis, and other types of pattern recognition. The human interface of NeuroShell was specifically designed for simulation of human experts, problem classification and diagnosis, credit analysis, forecasting, and other advisor services. In many types of problems for which IF/THEN rule based systems have been applied, NeuroShell may be more useful and far quicker to implement. NeuroShell includes two versions of the main program, called Binary and Analog. Binary is useful for "YES/NO" expert systems, while Analog requires the use of confidence factors along a scale. For example, zero might mean "NO", ten might mean "YES", five might mean "MAYBE", and nine might be interpreted as "ALMOST DEFINITELY". There are no built in limits to the number of questions or solutions. The maximum is limited only by available disk storage and memory. IN THE FUTURE ============= It is important to realize that while in a limited sense Neural Networks can be said to "think" for themselves, the resemblance to human thought is slight. The size of today's Neural Networks is comparable to the brain of a Honey Bee. They cannot duplicate the complexity, memory or reasoning power of the human brain yet, but they will eliminate a great deal of software development time, because they can discover solutions without conventinal step-by-step programming. NeuroComputers, the more powerful and expensive hardware implementations, are not at all "fuzzy" in their thinking or logic, and can perform higher mathamatical functions with precise accuracy. This will make possible many computer applications that were formerly too expensive or too hard to develop. Here are some examples: SPEAKER INDEPENDENT CONTINUOUS SPEECH RECOGNITION - Within three to five years, we will have an inexpensive working Voice Typewriter, which can be trained to recognize the voice of several users. ROBOTS - Intelligent trainable household robots will do all our unpleasant household chores, including the windows. You will train your PR (Personal Robot) by putting him through his paces once or twice with a hand-held remote control device. HANDWRITING RECOGNITION - already in use, will become commonplace, for use in Banks and other businesses. KNOWLEDGE PROCESSING - Researchers have already demonstrated "Instant Expert Systems" in the medical field, using Neural Networks to extract medical knowledge from case histories without additional programming. STOCK MARKET PREDICTION - The Stock Market, more than ever, will resemble an electronic gaming network, where the latest NeuroComputers battle each other for a king's fortune. INTELLIGENT ARTILLERY SHELL - Can distinguish between friend and foe by subtle sonic cues it has been trained to recognize, and if it lands too close to friendly forces, it does not explode. MORTGAGE ADVISOR - Banking Institutions have a vast quantity of data involving past experience with mortgages and Consumer Loans that could be used to train a Neural Network to make future Credit Risk decisions. The Neural Network represents a remarkable technical leap in the state-of-the art in computer intelligence. It is a new building block in the field of Artificial Intelligence, just as the neurons in our own brains are the building blocks of our thinking. The Neural Network has allowed computer scientists to create the first convincing likeness of the brain in areas such as memory, vision, and self-taught speech. One industry spokesman believes that Neural Networks, and their faster hardware implementation in Neurocomputers, will finally bring about the "Age of Artificial Intelligence", with consequences as far reaching as the industrial revolution. If this is true or not, we definitely have come one step closer to computers that can actually learn and think.