Horizons of Science: Thinking Machines (1959)

[Music] Our lives are being rapidly changed by the development of large computing machines used in business, industry, defense to automate factories and to perform scientific calculations. They can do arithmetic and straightforward mathematical computations with incredible speed. As a consequence, they can carry out calculations much too complex for a human being. They can do simple logic and when given facts can make elementary decisions. In this respect, they can already be said to replace the human brain. The question that arises is how far can this process go? To what extent can the human brain be replaced by a collection of vacuum tubes and transistors? Man's progress is largely a consequence of his wonderful restless imaginative brain, his ability to think intelligently in coping with the problems of his environment. The question is what is the nature of the thought process? What is it that enters into what we mean by thinking, by ideas, by imagination? If we can identify these processes, perhaps they can be built into machines. We haven't gotten very far yet, but in the short time that concentrated development on so-called thinking machines has been underway, enormous progress has been made. In this issue of horizons of science, we will see some of the researchers who are working on the problem using a variety of approaches of making machines duplicate the brain in the long run to make intelligent thinking machines. Mathematician Claude E. Shannon of the Massachusetts Institute of Technology designed this electronic maze not to prove any special point, but mostly for fun. It happens to show, however, one way in which a machine learns and acts on what it has learned. The movable walls of the maze can be arranged to make millions of different patterns through which the artificial mouse must travel to reach its goal, an artificial piece of cheese. To find its way through the pattern, the mouse proceeds as we would by trial and error. When it makes contact with the wall of the maze and discovers it cannot go straight ahead, it turns and keeps on turning till it has found its way again. It explores each step of the way with its copper whiskers. Here's a dead end. The only thing to do is go back again. Having been through the maze just once, the mechanical mouse is able to remember the correct path. set down anywhere along that path, it takes the most direct route without making a single false turn on through to the end of the trail. If Dr. Shannon puts the mouse down in a new part of the maze, one it has not traveled before, it again resorts to trial and error. But on re-entering familiar territory, it remembers every turn. Now the maze has changed again near its very end. Things go smoothly enough at first as the mouse proceeds on the basis of information it remembers. Now it's running into trouble. But fortunately, it is able to adapt to this new situation. It resorts once more to trial and error and acquires reliable new information to replace the old. [Music] But sometimes there's a problem without a solution. And even a mechanical mouse must have a helping hand. This agile three- wheeled little mouse contains a small bar magnet controlled from underneath the maze by an electromagnet that is mounted on a carriage driven by two electric motors. 100 relays and electrical circuits remember the directions for the mouse in much the same way that information is remembered in a dial telephone system. Dr. Shannon's mechanical mouse demonstrates that machines can learn by experience and can revise what they've learned when it no longer works. Two important characteristics at least of intelligent behavior. What can we learn about thinking from a game of chess? One challenging approach has been made by mathematician Alex Bernstein based on this ancient intellectually demanding game with its complex moves and its endlessly varying patterns of play. If chess had been played at the rate of a million games a second since the beginning of recorded time, only a small fraction of all the possible games would yet have been played. No human being can play a perfect game of chess, and neither can any conceivable machine. To find out how good a game of chess a machine might play, Mr. Bernstein and his collaborators prepared a chess playing program for the IBM 704, a digital computer that has performed 1 billion calculations in a single day in computing the orbit of an artificial satellite. The chess playing program is given to the 704 on a reel of magnetic tape. On the chessboard itself, the moves are made by Mr. Bernstein for both players. As he makes a move, he communicates it to the machine. The machine prints out the position of all the pieces, its own and its opponents, to correspond with the chessboard on every move. In calculating its moves, the machine considers the board square by square. Is the square occupied? By whose man? Is it under attack? Can it be defended? Can it be occupied? All this has taken a long time by computer standards. 1/10enth of a second. Now the computer proceeds to select its move. It has about 30 possible moves. After asking eight preliminary questions about each of them, it selects seven of the 30 for further analysis. It tests each of the seven through four moves ahead, considering its opponent's possible replies and its own possible counter responses in each case. It examines 2800 positions in 8 minutes. Now, the machine prints out its move. It elects to take the opponent's knight with its own bishop. Mr. Bernstein takes the machine's bishop with his queen. The move is recorded, but the machine rejects the move as illegal. The difficulty is in incorrect coding, which is corrected. The game continues with the machine playing methodically and tirelessly. It's never absent-minded and never makes an obvious blunder. In individual moves, it often plays like a master. In a complete game, it can defeat an inexperienced player, but can be outwitted by a good one. This game has gone up to the 21st move. Mr. Bernstein attacks strongly, threatening the machine's knight with his castle. He records the move. The machine's response is a useless pawn move. Its unprotected knight is lost to Mr. Bernstein's castle. The machine recognizes its position as hopeless and resigns. After losing a game, the machine will still make the same moves again and lose in the same way. Someday, though not soon, Mr. Bernstein feels a program may be designed that will enable the computer to profit by its own mistakes and improve its chess game on the basis of its experience against human opponents. The machine we see here can perform a function that was once only possible for living creatures to perform. It looks at a drawing of a circle, a triangle, a square, or a pentagon and makes a decision about the essential nature of each image. communicating it to us on the panel above. It is called a pattern recognizer and it was designed by Mr. Leon Harmon, an electrical engineer in the visual research department of the Bell Telephone Laboratories. [Music] In scanning these figures, the pattern recognizer looks for and tries to extract the essential feature. In this case, three-sidedness. Within limits, it can recognize any triangle, no matter what its size or position. The pattern recognizer calls this image square in whatever direction it lies. Whether it is large or small, the figure is still properly identified as a square. Separate objects can be distinguished from line drawings. The machine can recognize that three distinct objects have been presented rather than a three-sided figure or four objects. The same thing is true of five objects or six. Pattern recognizers have a definitely useful purpose. We might one day have systems to read handwriting, detect certain cells under a microscope, or recognize individual human faces. Can we expand our notions about intelligence to include machines? There is no simple yes or no answer. Mr. Harmon states his point of view. The behavior of some man-made machines is impressive. Computers and other devices recognize patterns, learn by trial and error, process speech and visual information in very useful ways, and play abstract games like chess. They perform elaborate computations, make decisions, store and handle vast amounts of information, simulate almost any other machine, even design other machines. But they can only do as they are told, executing complicated instructions carefully planned for them. They have no imagination. They cannot create or invent. and therefore they do not think in any real sense. Consequently, we can't call them brains. In structure, the differences between computers and brains are vast. The building blocks which make up the brain are extremely complex. This is a drawing of a single neuron. There can be thousands of individual connections to each one. The neuron is microscopically small. It is less than 1,000th the length of a typical computer element, a transistor, which typically has only a few connections leading in and out. The difference in complexity between nature's building blocks and our electronic ones illustrates only part of the vast difference between brains and computers. Our most complex computers have something like a million elements in them. While the brain contains roughly 10 billion. That is the human brain has 10,000 times as many working parts as the computer. And each one of them is many many times more complicated. Actually, no one really understands the intricate mechanism of the brain. In order to study the brain, we can investigate it directly. However, we can also work with models in order to test our ideas of how the brain may function. We can, for example, use simple electronic models of nerve cells. This device simulates some of the properties of a single living nerve cell. We call it an artificial neuron. We can connect numbers of these units together to get a variety of complicated responses to stimuli. If we connect a photo cell to the input, a series of pulses is generated somewhat like those that are found in the visual system and brain of living creatures. Increase the illumination and the frequency of pulsing increases. Decrease the illumination and the frequency of pulsing falls off. By carefully arranging and extending such simple models as this, it may be possible to gain more knowledge about how information is processed by living systems. We hope that in the evolution of our ideas about building machines to find ways of designing systems that will learn for themselves that may profit from their experience and do useful things in nonpredictable ways. It will take tremendous amounts of research and exploration to gain more knowledge about intelligent functions in both machines and brains. At some points, our findings from both areas may come together. Perhaps someday we will really be building intelligent machines. Also, a time may come when we will understand the complexities of the human brain. The roads to such understanding are long, but they are fascinating and tremendously challenging. Heat. Heat. [Music] [Applause] [Music] [Music] Hey,