Moravec, Hans;
Mind Children: The Future of Robot and Human Intelligence
Harvard University Press, 1988, 214 pages
ISBN 0674576160, 9780674576162
topics: | ai | robotics | brain | evolution
Predicts that robots will shortly be built at the human level. Unabashed high priest of the Strong AI standpoint. See also his extensive comments on [Penrose]. The excellent short movie on mammalian evolution in the mammal room of the American Museum of Natural History in NYC characterized evolution as : backbone - gives a diff way of propelling oneself (showed a fish), then Tetrapods - even better way (sea creature). To this one can extend Moravec's arguments (two-legged walking, elephant's trunk etc). I would really like to read his bio/brain sections, which are probably the more "surprise" containing parts. The predictions are probably off, but may not be much. Minsky had predicted a machine chess champ in 1980 - off by what looks like 20 years... Even if Moravec's timing is off - say it is 2200 - even then, the prospects are mindboggling enough. - AM --- A rich discussion on computers, viruses, artificial intelligence, robots, artificial life. At many places I found a discussion of his dreams rather than reality, in particular his prediction that human-like robots will be built within next 50 years i.e by 2038 A.D. We are nowhere close to that. The book truly points out that traditional AI researchers spent efforts on reasoning whereas the major problems are in perception and mobility. Discussion on repetitive behavior (pp.46-47): "Infinite patience would be an asset in a training session, but it could be exasperating in a robot in the field. In the cup fetching program I described earlier, you may have noted that if the robot finds the door closed and is unable to open it, it simply stands there and repeats "knock knock" without letup until someone opens the door for it. A robot that often behaved this way - and many present day robots do - would do poorly in human company. Interestingly it is possible to trick insects into such mindless repetition. Some wasps provide food for their hatching eggs by paralyzing caterpillars and depositing them in an underground burrow. The wasp normally digs a burrow and seals its entrance, then leaves to hunt for the caterpillar. Returning with a victim, she drops it outside the burrow, reopens the entrance and then drags it in. If however, an experimenter moves the caterpillar a short distance away while the wasp is busy at the opening, she retrieves her prey and again goes through the motions of opening the already open burrow. If while she is doing this, the experimenter moves the caterpillar away again, she repeats the whole performance. This cycle can apparently be repeated indefinitely, until either the wasp or the experimenter drops from exhaustion. A robot could be protected from such a fate by a module that detects repetitious behavior and generates a weak pain signal on each repetition, in the example, the door knocking would gradually become inhibited, freeing the robot for other pending tasks or inactivity. The robot will have acquired the ability to become bored." However this will kill desirable repetitions also. If the job of door knocking robot is to wake up a person sleeping inside, it is better that the robot does it till the guy inside wakes up, does the alarm clock not produce sound till we switch it off? I found the discussion on information theory interesting and am quoting it below, " Suppose a child's story begins with the words: Here's my cat. It has fur. It has claws.... Pretty boring right? Imagine now another story that now starts with: Here's my cat. It wears a hat. It totes a gun.. Better. The second story seems more interesting and informative because its later statements are less likely-cats usually have fur and claws but they rarely have fur and claws and they rarely carry hats and guns. In 1948 Claude Shannon of MIT formalized such observations in a mathematical system that came to be known as information theory. One of its key ideas is that the information content of a message goes up as its likelihood, as measured by recipient decreases (mathematically, as the negative logarithm of the probability). A series of messages has maximum information content when it is maximally surprising." (Then I think the information content of an infinite cyclic behavior must be zero. - Mali's comments) My measure of (Moravec's) effective computation works the same way. Each instruction executed by a machine is like a message. The more predictable its sequence of instructions, the less useful work a machine is doing. For instance, a program that causes a computer to simply add one to a memory location once every millionth of a second is doing almost nothing of consequence. The contents of the memory location at any time in the future are known in advance. But even the best programs are limited in how much " surprise " they can introduce into a computation at every step. Each instruction can specify only a finite number of different possible operations and choose from a finite number of memory locations, each itself containing only a finite number of possibilities. These sources of surprise can be combined using the formulas of information theory to express the maximum information content of a single computer instruction. - Amol D. Mali ( 6 August 1995)