Advanced Technologies Lead the Way to the Future of Educational Computing
Emily peeks at the display on her notepad-sized computer. Her professor has just sent her an urgent message to come prepared the next day for a field trip to study the ecology of a desert canyon.
She taps with a stylus on an icon representing her ecology class database and writes: “Get info on desert ecology.” Her handwritten letters linger for a while on the screen before the computer turns them into type. The computer displays the outline of the topics to be covered and the articles she should read. Other classmates have already left their thoughts about how they would like to divide the work.
In the library, Emily sits down next to a multimedia reader and inserts a CD-ROM disc prepared by the Public Broadcasting System. The narrator begins to explain the relationship between moisture, soil acidity, and the types of plants that grow in the desert. Emily does not know anything about soil acidity so she clicks the mouse to stop the video. “Define soil acidity,” she types. The computer not only provides a definition but also displays a graph of the relationship the documentary has been explaining.
In class the next day, the professor writes the plan for the day on the white board. The students listen intently instead of taking notes. The outline is automatically transmitted to their computers.
At the canyon, the students fan out on a cross-section, each clutching a computer, a walkie-talkie, and measuring instruments. The students enter temperature and soil acidity into a spreadsheet programmed ahead of time to handle just this type of data. Because the computers are linked via wireless modems, the information each student enters is automatically transmitted to all other computers. The combination of computer-to-computer communications, software for collaborative data analysis, and walkie-talkies allows the whole group to see where the project is progressing and to ask each other critical questions about what they are seeing.
In Emily’s next ecology class, the discussion turns to how insects walk. To explore this topic further, Emily puts on goggles with a monitor for each eye and a set of data gloves, and enters the world of virtual reality. She carefully studies the gait pattern of a simulated insect from a variety of viewpoints, then she reaches into the scene and constructs a simplified model of the insect from a construction set simulated by the computer. Next, Emily writes a program to control the movement of the simulated legs.
These scenes are no pure flights of the imagination. Many of the technologies discussed above are already in development, and a few, such as the wireless computer modems, are already being field tested. Although it’s difficult to predict the effect these advanced technologies will have on higher education, their impact will be significant. In this issue, we take a look at some of the directions the new technology will be taking over the next few years and beyond.
Smaller, Faster, Less Expensive
While it is often dangerous to forecast the future of a technology, two trends are clear as far as computers are concerned. The first is that computers have been getting smaller, faster, and cheaper for decades and show every sign of continuing to do so. This trend is starting to bring to fruition the benefits of technologies that have been on the drawing boards for decades, such as voice and pen interfaces and intelligent tutors.
A second trend is that computers are becoming ubiquitous. Cars, washing machines, and street signals are all getting their own built-in computers, and we don’t even notice. In the future there will be many more such computers, so specialized that we won’t even be aware we are using one. For example, companies such as Apple are planning not one multipurpose model for their pocket-sized personal digital assistants, but a whole set of computers with completely different uses. In this article Syllabus focuses on three key technologies: interfaces, connectivity, and intelligent applications. The material here is not meant to be all-inclusive, but merely to give a taste of what lies ahead.
The Changing User Interface
The user interface, the way we interact with computers, has come a long way since the time of punched cards and alphanumeric terminals. The graphical user interface, with its pull-down menus, windows, mouse, and trackball, will continue to evolve to become more responsive and intelligent. Other means of interacting with a computer—pen, voice, and data glove—promise another interface revolution just over the horizon.
The Pen-Based Interface
The pen-based interface, which allows a user to input data directly onto a computer screen using a stylus, has been around for several years. But within a year, a new generation of hand-held devices such as Apple’s Newton is expected to bring the interface into widespread use. With advances in more “intelligent” system software, pen-based computing is becoming more sophisticated. You won’t have to write each letter in a little square, as many pen-based computers force you to do, but you will have to pick up the stylus between letters—continuous handwriting is still too difficult for a computer to handle.
You will be able to edit text with your pen as well. Cross out a line of text on your screen, and it gets deleted. Circle a paragraph and draw a little insert sign at the place where you want to insert it, and it falls into place automatically. Elements of your sketches will be recognized, too. Draw a straight line or a circle freehand, and it will turn it into a line or a circle just as if you had drawn it in a drawing application.
As the interfaces develop, system software is getting smarter, too. In the future, your computer won’t have to be told explicitly what to do all the time. It will learn to understand your needs and become an assistant rather than a mere tool.
If you want to fax a letter you just wrote to a co-worker named Richard, you will simply type or write “Fax this letter to Richard.” Your computer will automatically look up his fax number in your address book, create a cover sheet, and place the fax in your fax out box, ready to go out whenever you connect your computer to the phone line. Other such assistants will automatically hunt for information in on-line services or check the availability of your co-workers to set up a meeting.
The computer will constantly watch what you are doing and look for ways to help you, based on what you have done in the past. If it misinterprets your intent, or you want it to do something differently, you will always be able to teach it to do the task in a different way.
The pen-based interface will probably first become popular on pocket-sized computers where you write and draw directly on top of the device’s flat LCD screen, but there is no reason why the same type of interface could not be used with a future PowerBook and other Macintosh models. Many designers and artists already use pens and tablets to input material into their Macintosh computers. One company, the Communication Intelligence Corp. of Redwood Shores, Calif., has already released MacHandwriter, which recognizes handwritten Japanese. It plans to release an English version in the near future.
Speech Recognition and Synthesis
In the next few years you will be able to talk to your computer, and it will talk back. You will be able to control applications and query databases, and your Macintosh will read your e-mail and faxes over the phone while you are away from the office.
The voice interface will be useful for pocket-sized computers whose screens limit the amount of material they can display. It will also be a plus in laboratories or at engineering stations where switching back and forth between the keyboard and the job at hand is impractical. It will help to proofread numbers in spreadsheets and may even improve the quality of student papers by reading what the students have written back to them.
You can already talk to the Macintosh with products such as the Voice Navigator from Articulate Systems, a combination of hardware and software that can be trained to understand hundreds of commands. You have to speak to the Voice Navigator one word at a time, however, and train it to understand each command by repeating it over and over again.
Even then, it may not understand you if your voice changes because you have a cold, are under stress, or just ran up a flight of stairs.
To overcome these limitations, Apple and other companies are working on speaker-independent systems that understand uninterrupted ordinary speech without having to be trained by the user.
A prototype speech recognition system called Casper can recognize hundreds of commands and even whole sentences. Casper does not break down if you have a cold, and even tolerates a foreign accent; one of the people demonstrating it around the world is French. There are limits, however. While Casper understands native speakers of American English about 95 percent of the time, a thick Russian accent would send the error rate through the roof, says Kai-Fu Lee, the principal scientist on the project.
Casper’s recognition consists of several steps. The commands you speak into a headset or microphone are processed with digital signal processing techniques to filter out much of the noise and unimportant variations in speech. Casper then matches the resulting patterns of numbers with those of some 160 different allophones—the sounds that make up the English language.
These in turn are compared to a stored dictionary of words and phrases. To make recognition possible, the search is constrained by rules about what constitutes a word—an “ahem” or a cough does not—and what sentences are grammatically acceptable.
Though Casper does not theoretically have a limit on the number of words it can be made to understand, in practice users of this type of system will be able to use a limited set of commands. The computer may understand, “Fax this file to John” but may not understand, “Fax the document to John.” You’ll probably end up with a list of commands next to your computer to help you remember just what the computer will understand.
The major drawback of Casper is the effort developers have to go through to integrate it into applications, Lee says. At present, companies have to hire someone with Lee’s experience to do the job.
The algorithms that make Casper possible have been used by researchers, including Lee, to build experimental systems for decades, but personal computers have not been fast enough to handle them until now. Casper now runs on Apple’s top-of-the-line 68040 microprocessor-based Quadras.
Even when faster processors and perhaps some special chips dedicated to speech recognition become available, there are limits to what a Casper-like system will be able to understand. Lee doubts that Casper will ever become an all-purpose dictation machine.
Getting a computer to understand speech the way humans do will require building in human-like intelligence, something few computer scientists even think about. Still, says Lee, even if they don’t make much progress on the machine intelligence front, there are plenty of speech-recognition applications to explore before this technology hits a brick wall.
Connectivity
Hooking up computers with cables or over phone wires is the most common type of network, good enough for sharing a printer, sending electronic mail, and passing word-processing files. To start sharing scanned color images over a network, you need more capacity.
An Ethernet network, which requires you to install a coaxial cable, is capable of transmitting information 10 times as fast as phone wiring. And although Ethernet is good enough for still color photographs, it is too slow to handle future multimedia applications such as video, sound, and animation, so many campuses are installing fiber optic networks, which are at least twice as fast as Ethernet. Wireless networks now under development promise to provide similar capabilities to Ethernet and fiber optic cables, without the expense of wiring a whole campus.
In July the computer industry won a big victory when the Federal Communications Commission, (FCC) the agency that oversees the nation’s airwaves, set aside a good chunk (some 20 MHz) of the radio spectrum for computer data communications.
Jim Lovette, Apple’s chief scientist for wireless communications, predicts that this will create a networking revolution in the next few years.
At present wireless networks are limited to sending information at about the same rate as a very fast modem or a fax (see illustration, facing page for comparison of networks). But with the FCC decision, that will quickly change. Within a year companies will be offering wireless modems that cost about the same as their telephone counterparts and they will be fast enough to transmit full-color images and digital video.
Apple is also exploring infrared as an alternative to radio, mainly for its future Newton family of pocket-sized electronic assistants. An infrared modem works something like the remote control for a VCR except that it does not have to be pointed at any one place to make the connection. Its signal bounces off the ceiling and walls.
Infrared radiation has the advantage of not going through walls, as do radio waves, so there is no possibility of interference and no FCC regulations to worry about.
Unfortunately, infrared does not work in the open air and even has problems in sunlit rooms. Lovette predicts that future wireless modems will probably include both radio and infrared options, switching automatically from one mode to another depending on which is appropriate.
Whatever the form, wireless technology will make it possible to set up impromptu local area networks anywhere. All students will need to do is bring their computers, turn them on, and click on the wireless icon on their screen. Each network will provide the information-carrying capacity equivalent of three to four Ethernet cables over an area some 50 yards in diameter, enough for even a large class to work together collaboratively.
Wireless will also put networking a whole campus within reach of most institutions. Instead of having to pay for installing fiber optic cables through old brick walls, educational institutions will just have to plug devices about the size of a lunch box into ordinary sockets at strategic locations around the campus. These devices will create on a small scale the equivalent of the cellular telephone network now provided by telephone companies.
The cellular telephone network is not suitable for transmitting data at high speeds, says Lovette. Not only is cellular costly, but data is much less forgiving of the constant fading in and out familiar to anyone who has used a car phone.
In practice, Lovette believes schools will find it practical to build a network combining wireless and fiber optic cable technologies, say, running a fiber optic cable from the library to a dorm or a classroom building and doing the rest with wireless.
Wireless networks also promise to be much easier to set up and administer than traditional networks. Putting in place all but the simplest of local area networks can quickly deteriorate into a black art, given the alphabet soup of mutually incompatible networking standards. Lovette believes the regulatory power of the FCC over the airwaves provides a unique opportunity for the computer industry to “do things the right way this time.”
Wireless networks will also provide the missing link most educational institutions need to connect to the nation’s high-speed networks. Congress last year provided $803 million in seed money to upgrade the Internet, the nation’s educational and scientific network. The plan calls for the creation of the National Research and Educational Network (NREN), a data superhighway that by 1996 should enable any educational institution, not just a select group of universities, to move large volumes of information, including digital sound and video.
With wireless networks, a high school or community college will need to install only one gateway to NREN for its students to be able to access the combined resources of the nation’s libraries, archives, and research centers.
The emergence of wireless and high-speed networking will augment the usefulness of collaborative software packages (see Syllabus 15) and make them less expensive, as well as more accessible to students.
The Coming of Intelligent Applications
The increasing power of the personal computer is making it possible to develop applications that are smarter and more responsive to the user. In the future, computers will not merely be passive tools used to put out newsletters or solve accounting and math problems. They will take an active role, giving students advice about things they may have forgotten as they work and correcting them after they are finished. Anyone who has used a spelling or grammar checker has experienced this type of application at a very rudimentary level.
Over-The-Shoulder Critics
Take a computer-aided design (CAD) package a student may use to design a house. Today these packages automate the traditional way of working with tracing paper and pencil but do not do anything for the thinking process that goes behind a design.
At the University of Colorado, Mark Gross, Casey Boyd, Andreas Girgensohn, and Gerhard Fisher are applying techniques invented in the 1960s called constraint programming. The idea is to have the design program keep track of the many constraints an individual designer faces.
In one prototype called Janus/ Constraints, different design elements in a house can be programmed with knowledge about their layout. Refrigerators and other appliances can be made to automatically snap to walls, counters to stretch to stay adjacent to an appliance when it is moved, sinks to stay centered underneath windows.
With another prototype named Janus, a designer lays out a kitchen while “critics” watch the designer work and make suggestions. If the designer tries to place the stove underneath the window, for example, a critic will warn that the curtains may catch fire. The designer can then explore the pros and cons of that particular design by reading the rationale behind the critic’s suggestion.
This type of software will be especially useful in group projects when the designer of one section of a project automatically creates constraints for other sections during the course of his work, Gross says, because it can identify possible conflicts early on and help designers resolve them.
Harvard’s Stephen Ervin is working on extending the idea of constraint systems to work across different media. Architects and designers use many kinds of representations including general maps of a geographical area, rough sketches, details, and cross sections. None of the software in existence lets them jump from one representation to another. The software Ervin is developing would let an architect’s rough sketch impose some constraints on the design when it is developed later in detail. A developer may look at a map of a region, decide to build a certain number of houses and then be warned, for example, that there is not sufficient sewage treatment to handle so many houses of that particular design.
Tutors
Another category of intelligent applications will be tutors. Dozens of tutors have been built over the past 20 years to teach everything from economics, physics, and programming to repair of helicopter rotors and operation of sugar mill recovery boilers.
The tutors use different learning approaches depending on the subject matter. Math tutors, for example, know an optimal way to solve a problem and help students when they stray off that path. These types of tutors have been found most effective with slow learners, says James Spohrer, Apple’s chief scientist for artificial intelligence in the Advanced Technology Group. Another kind of tutor simulates a particular problem such as a boiler going out of control and asks the student to come up with a solution.
Some tutors have been big hits: a Lisp tutor developed at Carnegie Mellon cut the time to teach Lisp programming by a third. Students working with an Air Force Electronics troubleshooting tutor gained in only 20 hours the proficiency equivalent to almost four years on the job. There is nothing magical about the results, explains Spohrer.
Troubleshooting in real life is very repetitive. Challenging or emergency situations occur only once in a while, so on-the-job learning is relatively slow. A simulation can present all the potential problems in a meaningful way and in a very compressed time frame.
The reason why there aren’t more tutors in classrooms, given their success, is twofold: many were programmed on Lisp workstations so they would not run on personal computers until recently, says Spohrer. More importantly, there simply haven’t been any good tools available to make it easy for educators to create tutors.
Tutors suffer the same kind of problem as other expert systems—transforming the knowledge of a teacher or an experienced boiler operator into the kinds of rules that make them up is a very time-consuming process often requiring the assistance of an experienced “knowledge engineer.”
Apple’s Advanced Technology Group is working to make the process fit for the rest of us. Working with Boeing, Spohrer has hit on the idea of representing the rules in an expert system as the kind of decision analysis tree that is already familiar to aircraft engineers.
To train mechanics to fix a pneumatic system, for example, educators would start by putting together a schematic model of the system on their screen out of a kit of components—simple icons or line drawings of valves, tubes, and compressors with all their important properties attached.
Spohrer foresees companies eventually offering kits—virtual hardware stores—of simulated components on a CD-ROM disc just as clip art companies now offer different sets of drawings. There may be a set for interior office design and one for chemistry lab apparatus.
Once the system is built, the expert uses it to simulate a particular problem, such as a valve getting shut, and then goes about solving it as he or she would in real life. The application automatically records the actions. When the expert has finished with one solution, some of the variables in the situation are altered and the process is repeated. When the actions to correct the problem diverge because of the new variables the expert has introduced, the system prompts the expert to explain the reasoning behind the new actions. In this way, the tutor generator application builds up the complex decision tree behind the simulation.
Beverly Woolf, a researcher at the University of Massachusetts, Amherst, who is one of the pioneers in the intelligent tutor field, and graduate student Tom Murray have already developed a Tutor Construction Kit, which includes aids to help a teacher visualize the relationships of different topics in the tutorial, and a special language to construct tutor-student interactions. The kit has been used by a science teacher inexperienced in programming to develop a six-hour, 200-unit course to teach statics. The teacher spent approximately 93 hours of preparation time per hour of instruction, which compares favorably with the 100 to 300 hours required for traditional tutoring systems, says Woolf.
Artificial Life, Neural Networks, and Genetic Algorithms
Traditional simulations, such as those described in Syllabus 18, focus on abstracting mathematical relationships in a model.
You enter the equations for predator and prey interaction into a package such as High Performance Systems’ Stella, and watch the sine wave graphs of their populations evolve on the screen. Researchers into what has become known as artificial life take a diametrically opposite approach. They start from the bottom up, modeling individual organisms with a few simple properties and watching complexity evolve.
A simple version of artificial life became well known in the early ‘70s as the “Game of Life.” You enter configurations of black squares in a spreadsheet-like grid, along with rules about their survival from generation to generation. If a black square is adjacent to one square or none at all, it dies; if it is next to two, it survives; an empty square adjacent to three squares gives birth; a cell surrounded by four squares dies of overcrowding. Run even such a simple program a few generations, and interesting patterns of shapes begin to emerge.
Maxis, an Orinda, California-based software publisher, has now put this type of simulation into several commercial products. In its popular SimCity, individual city blocks thrive or wither depending on such factors as crime, transportation, and urban decay. In the company’s SimAnt program, you can watch the emergent behavior of an ant colony you create.
Apple’s Advanced Technology Group (ATG) is working on Bugsy, a simulation of an ant colony that is integrated with data-analysis software so that elementary school students can start getting the knack of the scientific method.
Students can experiment with changing their ant colony in different ways. They can extend the range of an ant’s sensor for finding food, for example, or add beetles that prey on ants, and then examine graphically in Venn diagrams the resulting change over time in the numbers of ants and the health of the colony as a whole.
Maxis’ new program, called SimLife, takes the level of complexity of this type of simulation further. It lets you create entire ecosystems of creatures that can evolve over time. At the heart of SimLife is a genetic algorithm. Each creature has its own “genes,” strings of numbers representing different traits. When two individuals in SimLife mate, each gene splits into two, and the half-strings from the parents recombine to create a third data string. In this way, creatures you create in the artificial world evolve very much like those in the real biological one. Maxis has even added a random mutation rate, which you can turn up or down.
Genetic algorithms can do much more than simulate ecosystems. They can be used to evolve solutions to complex problems. In this case, the gene pool consists of strings representing possible approaches to solving a problem. Strings are weeded out according to which comes closer to providing a possible solution. The fittest strings are mated, and their offspring are weeded out again. The genetic algorithm has been used by David Goldberg at the University of Illinois, for example, to develop controls for a natural gas pipeline, a system too complex for traditional methods.
Neural Nets
Another exotic technology that has made its way to the Macintosh is neural nets. The idea is to use computers to simulate a network of neurons that, though they are far simpler than real gray matter, can be taught a few tricks. Just like real brains, neural nets do not have to be programmed. You simply present a neural net with a large enough body of data—say, mortgage loans or handwritten letters of the alphabet, and it learns to see patterns in the data.
Banks have been using neural nets to screen loan applications and to detect credit card fraud. Wall Street firms include them in their trading programs, too, though they are being tight-lipped about it. Some day neural nets may give computers vision. Dean Pomerleau at Carnegie Mellon has used a neural network for a vision system named ALVINN (Autonomous Land Vehicle In a Neural Network), which drives a van. ALVINN has learned to recognize a road well enough to have recently driven a 22-mile stretch of a highway at speeds of 55 miles per hour, Pomerleau says. Residents need not fear a careening van, however. A human driver is always behind the wheel, ready to intervene.
Half-a-dozen software companies are out with neural network software for the Macintosh with names like MacBrain, BrainMaker, Neurosoft, NeuralWorks, and Neuralist. Most of these products let you develop a neural network of some dozens of neurons on your screen in several layers.
The “synaptic” connections of each neuron receive both inhibitory and excitatory input from other neurons—represented by numbers, called weights. The knowledge of a network is not necessarily stored in a particular location but in the way the network is connected and the weights on the different synaptic connections.
Although these Macintosh tools are being used in universities for research ranging from neural anatomy and psychology to economics, they are not intended for the average Macintosh user. Most present you with a dizzying array of choices among the types of algorithms and neural architectures you may want to simulate.
Apple’s ATG is looking at neural networks for improving handwriting recognition. Michael Kaplan, the manager of the adaptive systems group, is training a neural net in his office to recognize handwriting, a process that at present requires the power of a Cray supercomputer. Apple is exploring the use of a special server developed by Adaptive Solutions Inc. of Beaverton, Ore., which is designed especially to simulate neural networks.
The Connected Network of Adaptive Processors System (CNAPS ) contains an array of four special chips with 64 processors each. These work in parallel to simulate the neural network to achieve 100 to 1,000 times the speed of a Cray.
One day such chips may make their way into a Macintosh or a Newton to recognize your handwriting or to enhance an intelligent tutoring system, but neural nets will probably remain largely an under-the-hood type of technology, just like their human netware counterparts.
