Three PC Programs Present Basics Of Neural Net Concepts

The past decade’s emerging neural network technology holds promise for a new breed of computers that may surpass conventional computers and even artificial intelligence (Al) technology in computing power and versatility. For solving many problems, scientists have dreamed of machines that work like human brains, and man-made neural networks come surprisingly close. Like the brain, neural networks are taught, not programmed, and, like humans, their responses are sometimes wrong. But in situations where a fast and accurate guess is needed, neural networks have no competition in the organic world. The most promising neurocomputers are dedicated machines, built from the ground up with silicon neurons and synapses. While the hardware technology is still in its infancy, scientists have had considerable success simulating such machines by running special software on conventional computers. Although the technology is probably another decade away from finding day-to-day use in laboratory research, simulated neural networks have already shown broad application in a number of disciplines, ranging from medicine to mathematics. As with the first computer languages, however, current neural network software is rarely useful right out of the box. The packages require training and often modification to suit a particular problem. Until recently, scientists hoping to use neural networks immediately have found little in the way of educational examples and tutorials to help them in this brave new world. Now three new inexpensive programs for IBM-compatible computers have arrived to provide scientists with simple but fascinating hands-on lessons in building and using “thinking” machines. BrainMaker, The Brain Simulator, and NeuroShell offer three different approaches to neural networks. Brain Simulator teaches how to build neural networks, BrainMaker teaches how they work, and NeuroShell demonstrates the techniques in action. The $99 Brain Simulator is essentially an educational toy—a set of simulated brain parts (neuron and synapses, represented on-screen by asterisks and lines) that can be assembled like Lego blocks. Yet from such basic components scientists can create surprisingly sophisticated neural networks. For example, using Brain Simulator’s basic tools, scientists can create pattern recognizers, boundary detectors, and even image enhancers—all quite complex networks— much more easily than by writing programs in a traditional computer language like Fortran to accomplish the same tasks. The program includes several sample networks that scientists can study and build upon. In one beginning lesson, scientists construct simulations of a computer chip’s familiar logic functions (AND, NAND, OR, and so forth) from a half-dozen on-screen neurons. More advanced lessons focus on concepts unique to neural systems like that of contents addressable memory (CAM). In the CAM lesson, the scientist draws a letter on the computer screen (a capital ‘A,’ for example) and the network searches for something like it among the patterns it has already been taught. The lesson is educational in more ways than one—the difficulty the network has in recognizing all but the most perfectly drawn letters will come as no surprise to neural computing scientists, but may be a quick return to reality to those novices hoping for a computer panacea. As a tutorial in a difficult field. Brain Simulator deserves a better manual than the slim walk-through supplied. Several key concepts are glossed over, and most of the networks’ workings are left to the student to fathom. The examples on-disk are excellent, but a clear explanation of what is going on would save students a good deal of laborious disassembly. The program is also hampered by its inability to self-correct, a scheme known as ‘tack propagation.” Researchers now believe that back propagation is essential to the creation of smarter and faster-learning neural networks. BrainMaker, on the other hand, does use back propagation. It also comes with a comprehensive 266-page manual and the excellent 222-page Introduction to Neural Networks. In this program, the neural “architecture,” or arrangement of neurons, is already in place, so scientists can focus more on experimenting with different inputs and outputs of the networks than on how to actually assemble them. Among the examples included are shape recognition, text-to-speech conversion (using written “phonemes,” or sound parts, as output to the screen), and basic artificial- intelligence decision making. Although BrainMaker will run on a PC, its internal process of teaching itself shapes, letters, or words can take several hours for sophisticated networks. An AT- or a 386-class machine can speed that up by a factor of 10 or more and is recommended. Much of the program’s appeal is in its visual depiction of the processes. Although the displays are limited to simple character-based graphics, watching the networks learn and evolve is half the education—and nothing short of amazing for neurocomputer newcomers. In one example, the network progressively makes better and better guesses at recognizing a shape or letter—in any orientation—that it is shown. At first try, all it sees are near-random blobs. But after viewing dozens of examples of the letter or shape, the network is able to correctly identify it nearly every time—even if it is upside-down in a far corner of its viewing window. Like The Brain Simulator, BrainMaiker works best with a Microsoft mouse. On a machine without one, BrainMaker’s use of drop-down menus and pop-up messages are of limited help. While all three programs are part of the no-frills first generation of neural network simulators, BrainMaker is the most restricted by the relatively crude working environment required to operate on a “plain vanilla” PC. After overcoming the interface, however, scientists will find BrainMaker a powerful and flexible simulator, with a broad range of experimental uses. At $99.95, it is also a bargain. At nearly twice that price, NeuroShell should offer something substantially more—and for scientists more interested in using neural networks than learning how they operate, it does. As a shell, it is designed to hide the underlying network, focusing on problem solving instead. NeuroShell’s forte is in identifying a solution based on incomplete, partially misleading, or unfamiliar information. While the program works hard at being educational (its 90-page manual is exemplary), it is primarily a working tooL As such, it has already found a niche in government and industry, replacing conventional rule-based artificial intelligence software in applications such as financial pattern prediction and decision making. If, for example, one substitutes animal “characteristics” with patient “symptoms,” the program can learn to become a medical diagnosis tool. Although scientists can learn much from this polished and impressive program, it might be wise to have a specific application in mind before investing the $195. A final (and much less expensive) product deserving mention is the well-received Explorations in Parallel Distributed Processing ($27.50, MIT Press, Cambridge, Mass). The 344-page primer includes two disks of eight ready-to-run neural network examples and, while basic, may be the best cheap introduction to the field yet. The user interface is nothing fancy, but the included C source code is a valuable asset for programmers.

Three PC Programs Present Basics Of Neural Net Concepts

Interested in reading more?

Become a Member of