Intro  Learning  Patterns  Real Life  Genetics  Classifiers  Biology  Neural Nets  Connectionism  Life  AI

Essays on Complex Systems Part 10 - Artificial Intelligence

Thoughts and ideas on Artificial Intelligence, how we think, etc.

[Bring over some of the sections from Talk.doc. Everything from here that belonged in Talk.doc has been moved there.]

Associative Memory and Thought

Examination of the components of a train of thought:

I saw a final baseball score on the news: 10 to 6 in 10 innings. First thought was that maybe a grand slam won it, but it probably didn't, because it would have been in the highlight tapes, but it wasn't. But that means that the 4 runs weren't all scored at once. If one winning run had been scored in the bottom of the 10th, the game would have ended, so you know the runs were scored in the top of the 10th, which also tells you whether it was the home or away team that won and where the game was played. The entire train of thought took maybe 5 seconds, and I'm not even a baseball or sports fan (so if any of the conclusions were wrong, that's why, but it's not important)...

Some components of the train of thought:

Free-association: 4 runs, oh maybe a grand slam? The strongest association comes to mind first. The free association itself is generative (in that it introduces imaginary new information that wasn't part of the given information) and is suggestive of a possible scenario.

Then the scenario is compared against all the given information to see if it violates any known rules for "how things work" in that situation, and accepted as a possibility, or rejected.

The free-association of imaginary information is difficult for an AI program, because although it's new imaginary information, it is not random -- it must relate generally to the scenario being considered and conform to its rules. I didn't consider the possibility that one team was already ahead, but a donkey ran onto the field and they had to stop the game, but what's to prevent a program from considering something like that?

It is not that those possibilities come to mind and are rejected: there is some process that prevents them from ever being formed in the first place. (It is probably the "probabilistic" nature of free-association. The most reasonable association comes to mind first, like a classifier rule that bids highest.)

Instead of wasting time generating millions of scenarios, something about the nature of the generative process makes it only possible to generate scenarios that make sense.

However, it's possible for that not to be the case, as in dreaming, so that's an interesting difference between the nature of the thought process when awake as opposed to when asleep. The imposition of the requirement that your thoughts make sense only happens when you're awake.

(Look for info on the development of General Problem Solver; what were people's descriptions of their own thoughts while they were trying to solve problems?)

Memory and Problem Solving

Examination of the process of recalling lost information

It took more than a day to remember the last name of someone I worked for 30 years ago. The last name was totally lost.

I started by running through the alphabet for first letters. Ruling out letters was easy. Became pretty sure it was N, M, or T. N was the strongest association, but no following vowels produced anything.

In another session, spent time on T, but nothing there either. "a" sounds, "o" sounds... nothing. At various times, when I wasn't thinking about the problem and then suddenly remembered that I could be, at that exact moment, something came to mind that seemed it should be it, but it was too weak and disappeared as soon as I noticed the feeling.

Strangely, in retrospect, none of those near-misses was actually very close. Also, I was convinced it was only 1 or 2 syllables, which it turned out not to be.

Finally went back to M, the weakest original association, but was thinking, "Hm... Nick M__, Nick M___, Mick M__. Hey, wait a minute. I know it wasn't Mick M__. Mick, Mick..."

The correct name came to mind almost immediately, as the result of this accidental Spoonerism which turned the first name into part of the correct last name. Was it a coincidence that earlier in the day I'd been listening to a comedy routine where all the first letters were reversed?

It's not often that you get such a good look at your own retrieval process, because anything that you can think of to try to remember is something you already know, and while you're trying to remember something, it's seldom that it occurs to you to examine the process itself while you do it.

The search was entirely sense-based, mostly auditory. That is, the criterion for whether a possibility rang a bell or not was based on whether the sound of a potential solution sounded right.

Also tried sight-based, remembering whether I'd ever seen an ad for the business in the local newspaper, what the layout might have been, and whether there was a name in the corner. And whether I'd ever seen a business card. And had a vague memory that when I'd looked up the name in the phone book once, it had been about midway into the book, at least not far toward the front or back. That actually relates as much to touch as to sight, not just the memory of the phone book looking about halfway open, but the feeling of it being balanced evenly. An amazing number of things go into a memory.

One thing that clearly is important is not just that memories and associations have different strengths, but that you can be aware of the strength at the same time as you have the memory. You don't just know that you have the memory. You know whether it is fuzzy or clear, and following that feeling is how you can feel your way to a solution.

This should not only be important to memory, but to problem solving, because solving a problem involves retrieving memories of how the problem, or a similar problem, was solved in the past.

The solution is easiest to find if you can find a memory of a similar previous problem, or if you can remember solutions to previous problems that would be useful, or parts of which would be useful, in the present. Either way, you're calling up memories and grading them by the strength of the response they produce. (Once again, as is often the case, graded response being the guide toward a solution.)

This ability to judge the completeness of a memory, the strength of its response, seems very important, and is a peculiar facet of self-awareness. How do you know? What are you judging it against? What's the standard? This is in a sense the direct feeling of how strongly a cell assembly is resonating. Or how strong its connections are, i.e. how often those paths have been traversed? In some way, it is a comparison of the current brain state with a previous similar one, and sensing how closely they match.

The second important thing is the ability to generate imaginary possible memories. That is, you know you can't get hold of the memory, but you are able to define the range the solution will fall into.

Armed with that knowledge, you can generate imaginary possibilities and judge whether they produce any response as a real memory, whether they ring a bell. If not, you move on to the next.

In the case of remembering the name, it obviously had to begin with a letter of the alphabet. That seems trivial, but it's not, considering the universe of possibilities that could be generated randomly, for example by a computer generating random possibilities. It actually narrows the search down a lot. (See the above section, on reasonable free associations.)

Any program whose aim is to solve problems must be able, in a similar way, to understand the context surrounding the problem enough to know what are the limits of the range the solution will fall into, so it can generate its imaginary possibilities in the most likely productive areas, as a person would.

Properly narrowing the limits of what the solution must be goes a long way towards actually finding it.

Was it Sherlock Holmes who remarked that sometimes a solution isn't so much found directly as by ruling out everything else until what you're left with is the only possibility?

Reasoning

Bongard problems and test questions of the type, "Which figure (or word) doesn't belong?" usually ask you to give the "best" answer. Solving the problem requires developing descriptions of the figures that categorize them in such a way that one of the designs doesn't fit.

This requires a type of reasoning that should be an ultimate challenge for any AI program. But aside from that, you create the categorizations based on experience (learning); the amazing thing is that there usually is consensus among people about which answer is best for most of the problems, even in situations where an alternative categorization can be created that supports a different answer with perfectly consistent logic.

Oddly enough, any answer could theoretically be supportable under the right circumstances.

Example: dog Manx Persian Siamese. The standard "best" answer is that dog doesn't belong because it's not a type of cat. However, to one person, Siamese might not belong because it is the only pet they have never owned. Nothing wrong with the logic there, and it is only assumed, never defined, what is meant by "best" on the test.

In a way, unless space is allowed on the test for describing the reasoning, this test would be invalid for testing reasoning ability, but would instead test how well a subject's socialization matches the "norm" (and how well their interpretation of "best" matches the norm).

This relates to problems of cultural bias in intelligence testing, except that here "culture" could apply to just a single person's experience.

Drawbacks of an AI program

As Hofstadter (Godel Escher Bach) (or someone quoted by him) points out, a truly intelligent program may be much different from what we ordinarily think, and not all of the differences between it and a more normal program of today are positive. (Hofstadter: An intelligent program, given two numbers to multiply, might think for 30 seconds and then give a wrong answer.)

Assuming that an intelligent program is one that learns from experience, it is unlikely that the knowledge it requires could be built in, due mainly to the enormous amount of knowledge it would need, just like us. Thus, it would need an enormous amount of experience, just like us. It might take 10 years or more (of constant supervision!) to train it.

(However, after the first one is trained, you could duplicate its learned knowledge in another unit, eliminating the need for that one's training.)

After training, its knowledge would be based on its experience, just like us, and its knowledge would therefore not just be factual.

Given a problem, its answer would be as likely as ours to be a mixture of fact and opinion, which might not be particularly desirable.

With a procedurally-trained program, you know that if you give it a particular problem, it will solve it the same way every time, giving the same results. This is very beneficial for something like an accounting application.

An intelligent accounting application might solve the same problem two different ways in two different runs, "deciding" the second time that it wanted to try a different technique.

This would present a particularly interesting situation to auditors or anyone in charge of monitoring the program's output.

In fact, the program would have to be subject to the same sort of oversight required of human accountants, since it would have the ability to make judgments that could be desirable or undesirable.

The oversight could be provided by other programs with similar training, creating a committee of programs. Who would be considered responsible for judgments or errors made by such program(s)!

In other words, in order to achieve the kind of judgment and flexibility that humans have, which is a principal aim of AI, you give up the kind of reliably deterministic intelligence that is one of the most useful things we rely on a computer to have.

Consciousness

Define various levels of consciousness. (See paper notes in Talk.cpp file for a good start at this.)

(Following describes a high level of consciousness.)

Consciousness is an entity's abstraction of its concept of itself, such that its conception of itself remains constant even while the entity changes.

To achieve that requires an even more basic ability: to be able to develop a conception of anything, which remains constant even while the thing itself changes.

In other words, to develop symbols.

(Seems like this must be a process where the brain's own internal activity can activate pathways that could otherwise only be activated by the presence of an external stimulus. When this is possible, the brain becomes an actively generative organ, rather than merely responsive. And imagination results. What is required for this? Circuitry loops where output from one process can serve as the input for another. This requires at a minimum sufficiently complex intertwined wiring.)

Note that you can imagine a very complex brain that is still only responsive. The threshold must be in the nature of the wiring, not just in the amount of it.

Try to develop a "Requirements for Consciousness" outline, like the "Requirements for Learning".

Some (Hofstadter?, find "Mind's I") have proposed that an intelligent system (read: complex adaptive system) will develop consciousness when it reaches a given level of complexity.

This seems unlikely. (Hofstadter says the same, in discussing neural nets?)

There are systems of great complexity in which we see no evidence of consciousness.

Individual people are conscious. But is a society of people? The individual people are incredibly complex, and conscious. A society composed of these people would seem therefore a more complex system than any of the individual people in it, and yet societies don't seem to have consciousness.

What makes this question very strange is that it is possible that the society is conscious, but it may be absolutely impossible for us ever to be aware of it, in the same way that our neurons almost certainly do not participate in any way in any sense of our own consciousness. They may be the substrate of it, without which it could not exist, and yet there is no way in which a neuron could have any awareness of the consciousness that it is a part of.

This would be a reason to create the "Requirements for Consciousness" outline, because it would give you a list of things to look for in a system. If you see the needed requirements, it would alert you that consciousness might exist there, and might be worth looking for.

There is a difference between experiencing consciousness and seeing evidence of it. That is, although we are probably unable to share in any awareness of our society's consciousness (if it exists), it might be possible to create yet another list, of things that conscious entities are or do that indicates that consciousness exists there (Signs of Consciousness).

Is it possible that consciousness is a cultural habit?

(Julian Jaynes, in Origin of Consciousness in the Breakdown of the Bicameral Mind argues that human consciousness is a very recent development, for physical biological reasons. But alternatively, could the reasons be cultural instead?)

It could be just a way of seeing things, of describing our (presumably) common human experience that has become almost universally accepted among humans (especially Western).

Have there been any societies in which, for example, there is no word or concept of "I" that distinguishes a person as apart from other people? That is, any society, ever, where a person did not experience him or herself to be something apart from other people, or the society, or nature, and where therefore that dichotomy did not arise in the language.

This is certainly an extreme of the Eastern satori state, which can be achieved intentionally, but have there ever been cultures where this state was the natural one?

I do believe that this could be the natural state of "unconscious" animals. They have sensations, and respond to the environment, but the sensations that the environment causes to happen to them, they perceive as happening within themselves, or rather, they simply perceive. There is no "within" them or "outside" them, and they do not feel themselves to be entities that do the perceiving. At the moment that the perception occurs, they are the perception itself. The sensations simply occur, and reactions result. They don't conceptualize themselves as the perceivers, as we do, at least partly because they don't conceptualize at all. Their experience is very different from ours. They don't experience themselves as an entity separate from their environment.

It seems likely that there never has been such a human culture (although some do minimize the "I" experience of the individual). Culture probably requires the capacity for abstract thought, and as soon as anyone or anything is capable of that, it seems likely that one of the first things they will abstract is themselves.

On the other hand, the existence of multiple personalities may tend to support the idea of consciousness as a habit. It is possible for one person to have more than one sense of self. Someone claims that Homer writes as though a sense of multiple selves was at that time commonplace (similar to Julian Jaynes claim). Sure, why not? Maybe that was the habit of the time.

-----

Although our stream of consciousness feels linear, and actually is linear, (and is usually in words), what if our brains are actually having thoughts all the time about various subjects, in parallel, and what we feel to be our stream of consciousness is just the one thought we're currently aware of, the one with precedence? Not just subconscious motivations, etc., but actual ongoing parallel thoughts occuring without our being aware of it. It's possible to have "blind sight", the practically usable capabilities of vision without the conscious awareness of having the capability of sight. It should be similarly possible to have thoughts, even coherent ones, that we're not aware of. Dreams and their seeming incoherence and fast topic changes could be the rising to the surface of those ongoing thoughts, while the focusing mechanism that normally makes one train of thought predominant is turned off.

-----

Sometimes the speed with which the brain processes and explains a sudden surprising event is so fast it seems impossible. You feel the explanation instantaneously, faster than you could analyze it sequentially and much faster than you can put the explanation into words. It seems to suggest that multiple processes receive the inputs and process them -- most likely independently -- in parallel, and the result is a "feeling" of explanation because the scenario generated is in concordance with another "feeling", the sense of how the world actually works. As mentioned somewhere else (likely by Hofstadter), you don't seem to think of innumerable potential explanations and discard illogical ones. You don't have the illogical thoughts in the first place. (Or, referring to the previous paragraph, maybe you do, in various centers of your brain, but your attention-focusing consciousness pays no attention to the illogical ones, and you're never aware of having had them, especially because they were occurring simultaneously and our consciousness can't deal with that. It would be like having multiple selves.)

Animal Consciousness

What animals are "unconscious"? It seems almost certain chimpanzees and gorillas have full self-consciousness.

I've long suspected that most dogs are not conscious, but that some might be (which would put dogs into a special borderline group worthy of study).

Knowing one's name isn't proof of being conscious. It is easy to form an association between a sound and good things that happen when you respond to it. It doesn't automatically mean you have an abstract concept of yourself. (Or that you are able to generate a symbol for anything. When you ask a dog if it wants to go for a walk, and it gets excited, is the excitement merely a conditioned response because that sound has frequently been associated with a good thing, or does the word cause mental images to flash through its mind of places that it likes to walk, etc.? Evidence of such images might be evidence of a level of symbolic thought.)

One interesting difference between how people respond to the world and how dogs seem to is that dogs (and cats, and others) seem more alert to, and immediately responsive to, the environment. Even when apparently asleep, they often respond to sounds that we filter out as inconsequential.

Dog and cat responses to things are noticeably immediate.

When a person, chimpanzee, gorilla, or dolphin responds to things, there is a momentary delay to every act. Their responses don't look immediate, even when they are quick. They look deliberate, sometimes even careful. Their manner conveys the impression that they have a sense that there are other ways to do what they are doing, and that they therefore must constantly monitor their own actions and decide whether to continue, must constantly decide how to continue.

What I suspect is that there is an extra level of processing that takes the extra time, and that extra level is the level at which the individual is not only aware of the stimulus, but is almost (but not quite) simultaneously aware that they are aware of the stimulus.

Responses of people (or conscious animals) usually look considered.

Sometimes you do see a dog appear as though trying to figure out what to do about something, which isn't exactly the same thing as this always-present delay, but might be related.

The extra delay in response time is small and subtle, but I believe it is important.

In situations where instinct (one level above physical reflex) is important, this delay is probably detrimental. But in other situations, the benefits of whatever processing takes place must make the delay worth the time. You could measure this delay and determine whether different animals do exhibit it to different degrees. Are some animals just slower-witted than others? If so, it may be because they have more to think about.

In order to think about something, you must be temporarily less responsive to external stimuli. Is the continual vigilance of a dog or cat additional evidence that it can't think about anything, and it has nothing else to do except exist in a continual state of responsiveness to immediate stimuli?

Consciousness also suffers from the same problem that intelligence does, in that its common definition, and the tests that are commonly proposed for determining when it exists, drag into the discussion factors that are really separate and not specific to consciousness.

For example, emotion, a biological arousal resulting from hormonal effects combined with a cognitive component, is impossible without the biological system. Yet an artificial system that is fully conscious but does not have the capacity to feel emotion is easy to imagine. So emotion isn't a factor in assessing whether consciousness is present, even though we feel intuitively that it has something to do with our sense of consciousness.

The common belief that the soul or spirit is something contained in the body, but separate from it, is quite understandable. Consciousness probably results from the nature of neural interactions, but can't be physically localized anywhere in the system. (It's an emergent property.)

In a sense, the body is the container of the soul, but more in the sense of being the medium in which it exists, and the working of which give rise to it.

However, consciousness in the absence of neurons (or some comparable physical medium) is similar to a hurricane in the absence of air. The potential for a hurricane exists only because of the rules governing how air particles interact.

Gathering together a bunch of air won't necessarily produce a hurricane. But without the air, you can't possibly have a hurricane.

Levels of Consciousness

Some very clever tests for possible consciousness of various types have been devised (described in a TV show -- Nova?). Reactions to seeing oneself in a mirror (treating it as another individual or using it for self-examination). Whether an animal follows the gaze of an experimenter (realizing that the experimenter must be looking at something). Whether a chimpanzee can tell whether a human spills a drink by accident or intentionally.

Zebras

One TV show said downed cows stop struggling after 10 minutes (even though, presumably, continued effort might solve the problem). One also sees on TV nature shows zebras that dash wildly from a lion, but, after it has caught them, almost immediately stop stuggling and sit quietly while the lion throttles them. Why? And does the situation offer any insight into the zebra's level of consciousness?

Their perception of the situation? The zebra is literally, instinctively, afraid of the smell and sight of a hunting lion, but unlike a human, not because of any extrapolation of what those stimuli imply or foretell.

They're not afraid of the pain because they don't know the lion will cause it until it does.

And they're not afraid of being eaten because they (probably) haven't learned that connection from seeing it happen to others.

They're literally afraid of the stimuli. (However, remember that even an octopus can learn by observing, so there could be much more cognitive component than I give them credit for.)

Why do they give up so fast?

1. Shock? With windpipe closed after the maximum exertion of the chase, this seems most likely.
2. It is an "uncovered situation", one to which no instinctual response has evolved because nothing has worked?

Misc.

As implemented in Robot.cpp, but maybe not mentioned much in my other programs, you must sometimes move away from a goal in order to reach it later. Very important, non-intuitive. Strangely enough, it may be easier for an entirely empirical system (such as a classifier) to develop these kinds of methods, since, having no conceptions, it can have no preconceptions.

One thing people do naturally is integrate input into a coherent structure, mold it into a "story line" that makes sense according to the guidelines imposed by our world view.

Simulating Physical Processes

Physical reality is a massively parallel computation performed on a virtually unlimited data set, both of which are impossible to achieve in any simulation.

Each part of each thing you try to simulate (for example, every atom) is essentially a small independent computer that is constantly "aware" of its surroundings in the sense that it is uninterruptedly affected by them, and constantly computing what to do next (the rules of the computation provided by physical laws).

For example, the motions of molecules in a gas are no particular computational problem for the molecules themselves. But it is a computation, made easy by the fact of massively parallel, massively decentralized, processing.

By contrast, a complete computer simulation of it is impossible, partly because of the huge amount of data required and partly because the method of sequentially processing one molecule at a time and moving on to the next is incapable of capturing the simultaneous nature of what really goes on.

When you choose not to extend your simulation below a certain level, you lose the ability to simulate anything that results from the interactions below that level.

Unfortunately, because nature is built from the bottom up, you may lose something critically important.

Simulation, though in many cases unavoidable, should be avoided wherever possible, especially because the simulation itself may consume more processing time than the thing you are trying to develop.

If a program is intended to interact with physical reality, it should be provided with the apparatus to actually do so, rather than giving it a simulated toy environment. As a particularly good example, see Genghis, below.

Much of what we include in the definitions of intelligence and mental process is so intertwined with and dependent on other aspects of our biology that it makes it almost necessary to include artificial life as part of artificial intelligence.

Examples of current state of artificial intelligence research

From Nightline, "Thinking Machines", March, 1997

Rodney Brooks at MIT has built a robot (COG) with TV vision, usable arms and hands, that is naturally explorative and learns about the world by its interactions with it.

Someone somewhere is trying to build an explicit database of "common sense" facts. Can ask the computer to find "a picture of a dangerous activity".

AT&T has an active AI development program. Voice synthesis, recognition and understanding, language translation.

Roger Shenk, Northwestern University, on natural language processing.

IBM's Deep Blue master chess playing supercomputer/program.

There do exist programs that do visual field comparisons such as I envisioned for Neural.cpp. One compares a picture of the user at a computer with a stored image of the person, for determining access to secure files. Obviously, the images won't be exact matches, so it must use various "how close are they?" comparison techniques. Fingerprint identification probably similar.

From Scientific American Frontiers TV show, "Robots Alive!", April 1997

[For each task, think about what are the most important problems that have to be solved.]

There is an annual AI Convention (1996 in Portland, Oregon):

1) Maze contest in a layout of office cubicles: start at director's office, find an empty conference room, then find offices of 2 professors, invite them to the conference room, return to director's office 1 min. before the meeting. Robots find doorways and determine whether there is activity in a room, usually by detecting motion via camera or sonar. Practical application of the techniques: smart wheelchairs. Contestants were:

• Carnegie Mellon Univ. (Amelia). Reid Simmons.
• SRI International (winner, multiple robots controlled by central computer). Kurt Konolige.
• Kansas State Univ. (Slick Willie).
• USC (Yoda).

2) Robots roam an area picking up tennis balls, including moving "squiggle" balls, and depositing them in a corner container. 10 min. time limit. Contestants:

• No. Carolina State (Lola) large, picks up balls singly in a pincher, hand-eye coordination.
• Univ. Of Minnesota (Walleye) small, traps balls with closable fence.
• Univ. Of Utah (Igor) large, slow, cages balls in hospital carrier with flaps that only allow entry.
• Univ. Of Bonn (Jeeves) blue plastic, brushes balls into container, reverses brush to flush out.
• Newton Research Labs (Vision Car). Color vision, and very lively chase routines. Traps and lifts balls (doesn't really grip) with grippers with rails on top that cradle the ball.
• Misc.: Iowa State's Cybot pours drinks and talks.

3) Carnegie Mellon Univ. (Pittsburgh, Chuck Thorpe, director. Dean Pomerlow, programmer), Martial Hebert, smart vehicle with cameras identifies obstacles, drives around them. Original inspiration was military. (The most advanced AI applications probably military, cruise missiles, etc.?) Also autonomous-driving auto. (Assistware Technology AutoTrack v.1.75) Uses any available information to determine probable lane markings, and keeps car centered. Warning or full-auto modes. Identifies obstacles ahead with radar. Program 10 years old (NavLab), previously used scanning laser.

4) Carnegie Mellon Univ. And MIT Leg Lab (Marc Raibert) machines that hop, do acrobatics, balancing by pogo-stick method.

5) Univ. of New Hampshire (Tom Miller) Toddler, is a top-heavy legged robot with motors and joints at same places as a human. Starting from scratch, it learns how to balance and walk on 2 legs from experience, by what works and what doesn't. (A genetic method seems likely.) Has balance and movement sensors: accelerometers, gyroscopes.

6) MIT (Rodney Brooks, worked on Mars rover program.)

1. "It", a robot that makes human-looking faces, can tell when something is too near (and responds), uses eyes and head to follow sounds, makes happy face when there's a lot of activity. Has microphones, eyes are motion detectors. Uses decentralized parallel multi-processing, like his bug machines (see Genghis below). Simple programs run simultaneously.
    
2. "Cog", his most ambitious project, a robot that learns about the world, including its own coordination, by trial and error. Has dual-camera eyes (wide range and center-field), learns to reach for objects held in front of it. Touch sensors on fingers. Analyzes wide angle view and identifies and turns toward motion. Also uses distributed multi-processing. (Such as module that turns neck to make head turn in same direction as eyes.) Also will turn head toward sound. Has variable-force arms and joints, adjustable depending on the task. Envisions, eventually, a robot with fully-human capabilities. Daniel Dennett, Tufts Univ. also working on this project.

7) Flying robot contest (at Georgia Tech, Atlanta). Autonomous flight, to pick up disks from within a ring, carry them over a fence, and deposit them in another ring. Contestants:

• Berlin, Germany: blimp
• Univ. Of Texas, Arlington: flying single propeller, "Texas Tailsitter", guided by off-board scanning lasers
• Univ. Of British Columbia: blimp, also guided off-board, TV cameras
• USC: helicopter
• Stanford Univ.: helicopter, uses GPS for guidance, by far the most capable machine, got and carried disk using a magnet on a string, just couldn't drop it.
• High school team: designed magnet in a cylinder to grab and drop disk

From Scientific American Frontiers TV show on brains, sometime in 2000:

Similar in appearance to some of the robots in the preceding section is one developed by Gerry Edelman at The Neurosciences Institute in San Diego (http://www.nsi.edu/).

It is operated by a supercomputer whose program appears to be a neural net program, with the network organized to match the module organization of the brain as closely as possible, and has many nodes (can't remember: 25k, 100k, 250k?).

It has a built-in "taste" for a particular characteristic (electrical resistance) of one type of block that it can encounter in its environment. It doesn't like other blocks. The blocks have different, consistent, markings.

Over time, it evolves methods to identify by sight which blocks will be the ones it likes.

So it must correlate its visual inputs with their implications. I.e. it must notice that the blocks it likes by taste also have distinctive visual markings, by which it can identify them from a distance. So although it has no explicit task, its implicit task is to "live the good life", learning about, identifying, and seeking things it likes, and avoiding things it doesn't.

One rule they use is "neurons that fire together wire together": you wire together neurons that are simultaneously active. That's how you choose how and where to create new connections. For possible practical use, see Complex.doc|Neural Nets.

Edelman also studies consciousness (using MRI, etc.), and has many good insights.

Examples of Genetic Methods Currently in Use

From Nightline, "Machines Like Us", August 1996

Computerized penguins for Batman movie. A somewhat trivial example conceptually. Each penguin was an object that had a repertoire of possible moves, based on a library of actual penguin motions. Each penguin chose its own mannerisms, and thus the penguins could appear as independent individuals when placed on the screen at once. (The penguins seemed to have the same weakness as the dinosaurs in Jurassic Park, the lack of inertia in their motions. A massive object doesn't move linearly. You can see the muscles strain before any motion results, then acceleration occurs, the motion peaks, and then deceleration occurs as other muscles slow the limb or body down.)

(Rodney Brooks, MIT) The mechanical robot insect, Genghis, that learns by trial and error (and probably genetic methods) how to move its six legs to walk. (apparently also discussed in an issue of Popular Mechanics). Gets from random to coordinated motion in about 2 minutes. Fascinating example because it is a physical robot, with sensors to determine whether its belly is on the floor and whether it is moving across the floor. Not a simulation! Real senses, real feedback.

Walking and limboing Luxo lamps and stick figures. (Univ. Of Toronto) Impressive merging of realistic computerized modeling of physical processes and the display of those models. (Try to find out what development tool they used. Seems unlikely they would have developed all their own software for it.) Uses genetic method to evolve the most efficient means of achieving the desired motion.

Advanced Investment Technology (pension fund manager). (Dean Barr, et. al.) Company evolved stock pickers and market timers that supposedly outperformed the S&P500 by 6%. The computerized timers can theoretically find non-intuitive correlations that humans wouldn't think to look for. Check on the continuing performance at a later date. Each evolved market timer tracks one stock. A very advanced use of genetic algorithm. Important question is what are the basic units of activity that the agents have in their repertoire? Presumably, their repertoire includes various fundamental ratios, etc., and apparently some technical data. Now, for example, can they only find combinations of the ratios they know how to calculate, or can they invent new ratios, formulas, and technical patterns to try out for predictive value?

Horse racing prediction, supposedly picked 38 of 50 winners. Another advanced example, since they must have addressed the same elemental-subunit problem as above.

Misc.

From Scientific American Frontiers: One artist has created a program with knowledge about people, plants, etc., which it then uses to create its own fairly good art with those as its subjects, choosing its own groupings, poses, etc.

Genetic method used to evolve spider web design rules to test whether good ones matched ones developed by real spiders. (SciAm Frontiers (Spiders) April 1999)

Used in various movies to create varied behavior for computer generated extras: penguins in Batman, people in Titanic.

Intelligences

From PBS TV show Discovering Psychology, of doubtful utility. Someone postulates 7 types of human intelligence, instead of one I.Q.: Linguistic, Spatial, Musical, Bodily Kinesthetic, Logical/Math, Interpersonal (understanding others), Intrapersonal (understanding self). These seem somewhat artibrary; you could invent a number of others. But the idea is certainly valid, and supports the idea that anything that involves any kind of information processing, that you can be good at or have an aptitude for, can be thought of as a type of intelligence.

It also raises the interesting question that if you don't know what intelligence is, how do you know what artificial intelligence is?

One of the pervasive and probably erroneous requirements for "intelligence" has been that it must be mysterious.

Much of what used to be considered AI isn't even considered AI anymore (edge and shape detection from video images, e.g.). They're just difficult tasks for which good methods were eventually found. As soon as we figure out how to do something, it loses its status as an artificial intelligence task. (This is what Hofstadter, Godel Escher Bach: p.601, attributes to Larry Tesler, and calls Tesler's Theorem: "AI is whatever hasn't been done yet.")

Misc

A British artist wrote a program to draw art. Program has knowledge about how various potential subjects work (people, trees, backgrounds), but chooses its own subjects, designs its own drawings, and drives a robot that physically renders its drawings to paper, including filling its paint buckets and washing them when it's done. (SciAm Frontiers, 1998?). (This is probably same as described above, in a different show.)

Cycorp, CEO Doug Lenat, is developing a program with common sense, world knowledge, that analyzes text to make sense of it, asking questions about what it doesn't understand. About 500-person-years in development.

MIT AI lab, Rodney Brooks now director, seems involved in just about every interesting aspect of AI I can think of.

Animal intelligence

From a show on dolphins:

Researchers created fake dolphin squeaks and associated them with items in the dolphin's tanks. When playing with the items, the dolphins made the same associated squeaks.

The people concluded that dolphins may use words in their communications, though that doesn't seem necessarily to follow from the experiment. It does seem to indicate that they are able to deal with words when trained on them. (Maybe they were just saying, "This is what those stupid humans call a [squeak]".)

In another experiment, dolphins successfully answered quizzes on items in their tank, in a sign language in which word order was significant.

A much more interesting experiment was that of a TV monitor installed in the dolphin tank where they could view it. They watched with much interest a picture of their trainer feeding the dolphins in another tank. When she picked up the fish buckets and walked away from the other tank, the dolphins watching the TV picture left the monitor and went to the place where they were fed.

In other words, the dolphins could not only make sense of the TV picture (which it's not clear all animals can do), but knew what it was, and understood that they were watching activities that were happening at that moment somewhere else.

They must have had other cues to know that those activities were taking place at that time. It would have been interesting to show them a taped version of the same thing. (It was interesting that whenever the trainer threw a fish, on the TV screen, the dolphins watching the screen lunged slightly towards where the fish would go if it popped off the screen.)

A National Geographic show described dolphins discovering the abstract concept of a new trick. When they stopped receiving rewards for all the tricks they knew, they eventually figured out that they had to invent new tricks, which they then did in great quantity.

From PrimeTime Live On Animal Intelligence

Alex, a parrot, has been trained to answer quizzes about objects: type of matter, colors, shapes, and even expresses other ideas appropriate to context: "I'm sorry", "Wanna go eat dinner". Also described in National Geographic show, and others.

Animal languages and dialects?

Whenever we unite individuals of any other species, we start with an assumption that they'll get along or understand each other, as though their communications are standardized across the species.

Couldn't local languages exist? Wolf cries in one valley might relate only to things peculiar to events in that valley. A group of dolphins or whales might have sounds unique to it. Intelligent animals drawn from far apart might not be able to communicate, except very simply.

In another show, on gorillas (or chimpanzees?) that had been taught sign language, one was shown lounging around playing with a toy cow and signing to itself, "Cow, black, black, black, cow, stupid, stupid, stupid cow..."

Why would it have to sign to itself? Couldn't it just think the concepts? Maybe not.

You can't have symbolic thought until you have symbols, which must consist of something. In our case, the symbols have several forms: textual, auditory, and others such as sign language.

It is possible for us to think quietly and invisibly mainly because it is possible for us to "hear" the sounds in our heads without uttering them. It is difficult to have coherent thoughts without resorting to this internal voice.

For the gorilla, the only symbolic units available are the hand signs. Maybe signing to itself is the equivalent of our thinking to ourselves, and it can't think coherent logical thoughts without actually making the signs because those are the only units of meaning it has available to manipulate.

Could it hear the sounds if it has learned them, even though it can't make them? I doubt it, because the internal voice is usually a person's own voice, and one does not just hear it. Part of the activity is the imagining of making the sounds.

Perhaps a gorilla could imagine one of its human friends making the sounds, but again I doubt it. However, it should be able to imagine making the hand signs without actually making them. Another question is why did it think the cow was stupid? And did it understand that the toy cow was only a representation of a real cow?

From Scientific American Frontiers:

An octopus that previously had never been able to open a stoppered jar to get at a crab inside was able to do so immediately after watching another octopus do it. (This was an absolutely amazing demonstration; you have to see it to believe it.)

From a different show:

Several cephalopods (octopus, cuttlefish, squid) have large brains and good eyes. They can change their own color and pattern to match environment. Automatically?, or intentionally by analyzing what's there and copying it?

Some experimenters found their cuttlefish in captivity to be friendly and playful.

Animal Einsteins: Scientific American Frontiers January 1999

Very clever experiments in which animals exhibit greater world knowledge and/or reasoning ability than one would expect:

Alex the talking parrot has appeared in numerous TV shows on animal intelligence.

Sea lion learning categories of symbols (numbers, letters, etc.) = reasoning without using language.

Chimp, Sheeba: experiment hiding objects in small model room of a matching real room, ability level about the same as human 3 year old.

Rhesus monkeys in "3 > 2" experiments: random untrained monkeys almost always choose the box containing 3 apple chunks.

"Looking time" measurements to judge ability to add: monkeys surprised when one object is added to a box with one already in it, but when opened still has only one in it. (Contradicts world knowledge.)

Sheeba also can add, using fruit (objects) or numbers, and fractions.

Ravens learning to reel in meat tied on the end of a string, gradually refining their method by what works.

Tamarind monkeys, expecting gravity to drop a Fruit Loop straight down through a tube, can't learn that it doesn't until trick apparatus is laid on its side, so expectation of gravity is removed.

Chimps using sticks to get honey (like termites in wild).

Tamarind presented with candy placed in or next to various tools (hooks, sliding trays, etc.) for obtaining it always chooses the tool that works best, on the first try, i.e. not learning which tool works best, but assessing the relevant features by instinct.

Chimp only warning another chimp of danger when circumstances imply the other chimp doesn't already know, i.e. awareness of what facts another individual knows. They do better than 2 1/2 year old child at similar test. Similar experiments with tamarinds: apples hidden in boxes while second experimenter either witnesses it or not. Monkey very surprised when (and only when) human does something he should know better than to do = monkey knows what facts the human should know.

Misc.

On the other hand, if dolphins and whales are so smart, and so communicative, why haven't they learned to stay away from tuna nets and whalers? (Answer: because it is the tuna that follow them!)