What is AI?
To answer this question, we first have to answer:
What is/isn't intelligent?and What is/isn't artificial.
1.1. What is/isn't intelligent?
We have a few choices here:
(i) Everything task that exceeds a certain level of complexity requires intelligence to perform. So any system that can perform tasks of that level of complexity is intelligent.The problem with this view is specifying that level of complexity, and saying what makes it special -- special enough so that everything below it can be called "unintelligent" and everything about it can be called "intelligent."
A second choice:
(ii) Whatever normally requires intelligence (say human intelligence) to do, if it is done by any system, is intelligent.For example, doing arithmetic normally requires intelligence. So anything that can do arithmetic is intelligent.
The problem with this view (if it is a problem) is that it defines intelligence in terms of performance:"intelligence is as intelligence does," and then it grounds it in human (or animal) "intelligence": "whatever can do what humans (animals) can do is intelligent."
But not everything humans (animals) do is intelligent. So we would first have to say what human performances are intelligent (and perhaps even why, and how) and then say that whenever those are performed, by any system, they are intelligent. (Which performances are those then, if walking is not one of them but doing arithmetic is?)
A third choice:
(iii) Only things that normally require human intelligence to do and are done in a certain way are intelligent.The problem with this view is that you have to say what that "certain way" is.
Here are some candidates:
To be done intelligently, something must be:
(a) done computationally (which computations, then? any computations? if only some computations, which? and why those?) or
(b) done consciously (why? and how do we know it's done consciously?) or
(c) done by a conscious system (why does it have to be conscious?)
Let's see if we do better with "What Is/Isn't Artificial?"
(i) Biological systems are natural; machines are artificial.So a rabbit is natural and a toaster is artificial: But what if a toaster (or a computer) grew on trees? Would that make that very same system into another kind of a system, a natural one, even though it was functionally and structurally identical in all respects? (What kind of a difference is that!?)
Or if we succeeded in synthesising a rabbit, molecule by molecule: would that suddenly make an identical rabbit artificial?
So what is a "machine" anyway (apart from the fact that it happens to be man-made)?
(ii) A virtual system in a computer (e.g., a computer simulation of a chemical reaction, or a virtual-world simulation of a visible object) is artificial, whereas a chemical reaction or a physical object is real.What is a robot, then? It is not a virtual object. Does that mean it is not artificial? Why does something have to be computational to be artificial? What is computation, anyway, and what is special about it?
(i) Computation is what mathematicians do when they "compute."This sounds circular, but actually it's just the performance definition again: It's whatever mathematicians are doing when they do what they do. The rest is about trying to say what it is that they are actually doing when they are computing:
(ii-a) What mathematicians are doing when they compute is captured by the formal notion of (a) the Universal Turing Machine, (b) General Recursive Functions, (c) the Lambda Calculus, (d) Post/Kleene Machines (etc.), all of which turn out to be formally equivalent to one another. That is what computation is, and that is what mathematicians do. Anything mathematicians do will always be captured by these (equivalent) formal notions.The above is also called the "Church-Turing Thesis," and it simply says that everything mathematicians have ever intuitively and practically meant by "computing" can be done by, say, a Turing Machine. So computation is what a Turing Machine does.
What does a Turing machine do?
(ii-b) Computation (i.e., what a Turing Machine does) is symbol-manipulation: formal symbols, arbitrary in shape (e.g., "0", "1") are "manipulated" (i.e., combined, recombined, written, erased, re-arranged) on the basis of formal rulesFor example, the symbols can be interpreted as quantities, standing for salary payrolls, or as the outcomes of scientific experiments, or as numerical calculations; they can even be interpreted as words and statements, true statements, about the world.
(algorithms, syntax) that operate on the symbols' shapes (e.g. "if you see a '0,' erase it and replace it by a '1'"), not on their meanings. Yet, if you have found the right symbols and manipulation-rules (the right algorithms), you can do remarkable things with them, so much so that the input symbols and the symbol-manipulations and the resulting output symbols will all be meaningfully interpretable.
One important feature of computation is that the symbol-shapes (the notational system) doesn't matter: it is the rules and manipulations (algorithms) that matter, not the notation they happen to be formulated in. You could have used a completely different notational system to do exactly the same computation.
This is also the basis of the software/hardware distinction. It is the programme that matters for a computation, not the hardware details. (The programme of course has to be implemented on some hardware, even if it's just through someone doing the calculations by hand, but the computation itself is independent of those hardware details: radically different hardwares could have done exactly the same computation: the only thing they would all have in common would be the programme.)
This does give us a good definition of what is and is
not computation: Anything that is what it is purely because it is executing
a certain symbol-manipulating algorithm, and not because it is a certain
physical system (obeying a set of differential equations), is computation.
Anything else -- anything more or less than this -- is not computation
(or not just computation)
Sample Question: All nontrivial computer programmes do something that is "intelligent." What distinguishes AI from the rest of computer science? Discuss principles and give examples.
2.1 IQ tests measure itThe concept of intelligence -- the intuitive idea and the everyday observation -- that some people are "smarter" than others -- is old. The idea that you can measure how intelligent people are is newer. Intelligence tests are designed to give higher scores to those people we consider more intelligent and lower scores to those we consider less intelligent.
To design such tests, we of course have to have some prior way of telling who is more and who is less intelligent. Than whatever that prior way is -- let us call it the "criterion" -- the questions in the test are picked so that those who can answer more of them correctly are more intelligent, according to the criterion. An example of a criterion might be how well people do in schoolwork, or in later work in life; or, for children, it might be that, say, a 9-year-old who can do the kinds of things the average 11-year-old can do is smarter than a 9-year-old who can only do the kinds of things a 7-year-old can do (the so-called "IQ" or "Intelligence Quotient, the ratio of the mental age to the real age). (There are of course problems with all these criteria, but at least they are criteria, and they can be used to pick out which questions we want to include in the test, because they correlate positively with the criterion, and which we want to throw out, because they do not correlate with the criterion.)
But the trouble with intelligence-testing is that it tells you who has more of it and who has less of it, but it does not tell you what intelligence itself is. For more about IQ and the factors underlying it, see: Arthur R. Jensen (1999) Precis of :. Psycoloquy: 10(023) Intelligence g Factor (1)
We don't know what intelligence itself is, but we do know it when we see it (and we can pretty much tell when there is more of it or less of it): How? By what it can and can't do. In general, the one we consider smarter is the one who can do more (of a certain kind of thing: see the Jensen paper for the difference between general and specific skills).
2.2 Intelligence is as intelligence does
But IQ tests are just about intelligence differences. What about just plain intelligence itself? What about what it is that all normal people share, whether they have higher IQ or lower? Let's call that generic human intelligence -- and whatever it is that makes all normal people able to do the kinds of things all normal people can do, that's intelligence.
So the question then becomes "what is it that makes a system able to do the kinds of things normal people can do?" AI is meant to provide the answer to that question. Let us call this the "How?" question.
Individual differences in intelligence -- the kinds of things some people can do and others can't, or some people can do better than others -- might possibly give us clues about the answer to AI's How? question: Maybe there are "modules" in intelligence, so that it consists of a set of independent abilities (but Jensen's work suggests otherwise; that although there do exists some specific skills -- musical, spatial, mathematical, verbal -- most of intelligence co-varies as one general "g" factor).
2.3 Individual differences
Yet it seems clear that no matter how long or hard we measure differences in intelligence, that will not in itself answer the How? question.
Besides individual differences between human beings in what they can and cannot do, there are also differences between species -- and not just sensory differences (some species see things better than we can, or even have different senses, such as sonar) or motor differences (some species can fly and swim) but also "cognitive" differences (memory, spatial analysis, etc.).
2.4 Species differences
But it is not the species that can do more than we can do that might be relevant here, but the species that can do less: After all, out abilities evolved out of their abilities. Maybe we should try to model animal intelligence (animal AI) first, before trying the harder task of modelling human intelligence?
In some ways this might be easier, except for one possible problem: No other species than our own seems to have language, and that seems to be at the heart of our own intelligence.
So after all, perhaps we have no choice but to take on the task of modelling generic human ability (Jensen's "g" factor). But we clearly cannot start at the top. We have to start with "toy" fragments of our total ability (as AI has done, with chess-playing, scene analysis, problem-solving) and then try to "scale up" to our total generic capacity.
2.5 Generic intelligence (g)
It is some of the dead ends that AI has run into in trying
to scale up to our total generic capacity that will prove to be illuminating,
in trying to work out what the best path for AI will be.
Sample Question: What questions can AI answer that IQ testing cannot, and how? Discuss concepts with examples.
There really is a difference between these two forms of engineering intelligence (one forward and one reverse), but we still have to ask what a "machine" is? Because if we cannot say what a machine is, then there is no difference between the two kinds of AI, except in motivation.
With natural things, we often do not know how they work, so we suppose they are some other kind of thing. But don't natural things have mechanisms too?
So when we raise the How? question about intelligence, all we are asking for is a causal explanation of how the systems we call "intelligent" (such as ourselves) are able to do what they are able to do. The answer to the How? question is a causal system that can do what they can do.
(Obviously, we have to understand the system, e.g., because we designed it. Just pointing to another system as an explanation, because it can do everything we can do, is not an explanation unless we understand how that other system does it! So building one while sleep-walking isn't enough: we need to understand the causal processes involved.)
Harnad, S. (1994) Computation Is Just Interpretable Symbol Manipulation: Cognition Isn't. Special Issue on "What Is Computation" Minds and Machines 4:379-390
The important thing to remember is that the symbols are manipulated on the basis of their shapes, not their meanings. The shapes of symbols are arbitrary, and they can be taken to mean anything. (There is no resemblance or causal connection between the shape of the symbol "apple" and those round red things that it refers to.)
The important thing is again that these rules apply mechanically, that is mindlessly, and are only formal, being based on the shapes of the symbols, not what they mean.
To put it another way: Every implemented symbol system is a dynamical system, but its dynamics are irrelevant to the computation it is performing:
Examples of implementation-independent symbol systems are arithmetic, logic, chess and language. Examples of implementation-dependent dynamical systems are motion, heat, sensorimotor transduction, and turbulence.
This is just to remind you that not everything is computation
(i.e., sometimes the phsyical dynamics are central to what a system is
Sample Question: What Is and Is not Computation? Say why and how, with examples.
The essence of Turing's insight is: We are not mind-readers, not even with one another. The only way you can tell that anyone else has a mind (intelligence) is by what they do; you cannot go into their minds and make sure. So if any system can do what we can do, it has a mind too.
This version of the TT has some problems:
(1) The penpal TT is not total. We have many more capacities besides our penpal capacities (although our penpal capacities are pretty central, being based on language).
(2) The penpal TT could in principle be passed by computation
alone (a symbol system).
If so, then it is open to a famous counter-argument by Searle: Since a computer programme is implementation-independent, Searle could himself execute all the code for passing the TT without understanding a word of what his penpal was talking about (e.g., if the TT was conducted in Chinese). This is not what we mean by intelligence.
(3) Combining (1) and (2): If you included a photograph (or any other physical object) with your letter to your penpal, the symbol system could not discuss it with you, as a normal penpal could (unless you also described the photo in words). This would easily be detected as failing the Turing Test. Maybe this is why Searle's argument works too: Because implementation-independent symbol manipulation is not enough to pass the TT: The TT-passing system needs some noncomputational (dynamical) capacities too, especially sensorimotor capacity.
6.1. ComputationalismSee Harnad, S. (2001) for a discussion of what's wrong and right About Searle's Chinese Room Argument. "Weak AI" is just using computers to try to model anything, including intelligence. "Strong AI" (or "computationalism") is the theory that intelligence is computation, and can be implemented, not just simulated, by computation alone.
Module 5 explained what computation is: It is (1) implementation-independent, (2) syntactic rule-based, (3) semantically interpretable, (4) symbol manipulation.
Symbols (4) are just arbitrary objects. It doesn't matter what the object, or its "shape" is, because any other object could have been used. The choice of object is just a convention we agree to use, a shared notational system, like agreeing to speak English or Chinese. The "shape" of the word we use to stand for something has nothing to do with its meaning. (What a word such as WORD looks or sounds like has nothing to do with what it means -- it means what we mean by "word," just as RED means what we mean by "red." The words do not "resemble," nor are they physically connected to what they mean in any way.)
Implementation-independence (1) is related to the arbitrariness of the shape of the symbols we choose to use. A computation is the same computation no matter what programming language you write it in. It is also the same computation no matter what hardware you run it on.
The symbol manipulations are based on "syntactic rules" (algorithms, programmes) (2) which operate only on the shapes of the symbols, not on their meanings. The best thing to keep in mind here is a Turing Machine: A "0" [or any other arbitrary symbol] appears on its reading head. The machine is in a certain "state" at the time, and let us say that the state [which is the implementation of the rule] is the following: "If you read a "0" while in this state, erase the "0", write a "1" and move to the next symbol on your reader."
That's symbol manipulation, based on symbol-shape, and not on symbol meaning.
6.2 Searle's Chinese Room
In some ways, there is a split between symbolic and robotic AI, and it is along the lines of the split between the symbolic (penpal) and robotic TT. Symbol systems are good at some things (calculation, reasoning, problem solving), but not so good at others (sensorimotor activity, learning. Hybrid systems seem to be the optimal choice.
6.3 Symbolic AI and Robotic AI
Another natural split is between skills and "knowledge": Skills tend to be procedural and sensorimotor (i.e., robotic) whereas knowledge tends to be proposition, factual (i.e. symbolic).
6.4 Know-how and Know-that
Consider the photo you enclose with your letter to your penpal: A picture is worth not only 1000 words (symbols) but an infinite number. You can describe faces in words will doomsday, it still won't substitute for the sensorimotor skill of being able to see and recognize them.
And what's in a name (of anything) if it is not grounded in that sensorimotor know-how?
An intimate part of knowledge is knowledge-acquisition, and that too is a skill: We learn to recognise faces through sensorimotor experience. Moreover, learning is an essential part of the robotic TT: Many of our sensorimotor skills are learnt rather than inborn, and the most fundamental of them all is categorisation: Our capacity to learn to sort and label the things in the world. Any system that could not do that would fail the TT from the outset.
6.5 Knowledge, Learning and TT-capacity
Our brains are remarkably powerful at induction. No man-made system yet comes close to the learning capacities of the human brain, but as we scale up toward the full robotic TT, we will have to model this powerful capacity.
Note that mathematics is deductive, whereas science is inductive: What then, is AI? Is AI an experimental science, to find out what can and cannot be done with certain symbol systems, or is it a mathematical one, to prove what can and cannot be done with certain symbol systems? Is an AI programme more like an experiment or a proof? What about a robot? Is reverse engineering more inductive than forward engineering? Is real robotics more inductive than virtual robotics?
Inside the child's brain, a learning mechanism is learning to detect the right features and apply the right rules. Supposing the child does well for a while, then it gets one wrong: What has it done wrong? Which of its features and rules does it need to change? This is the "blame assignment problem." If instead, the child is starts doing well, which of its features and rules should get the credit.
Sometimes the child will overgeneralize a rule: "If it's brown and furry, it's a bear" (this is not necessarily a conscious rule). Then it's shown a polar bear and gets it wrong: back to the drawing board, but what was wrong?
With only a few features this is not so hard, but where there is a huge number of potential features and rules, this problem can become a very hard one. And it is the problem faced by any system that hopes to scale up to human-scale learning capacity.
This is an instance of the frame problem. It is usually described as the problem of knowing what does and does not stay constant (invariant) when there is a change. The solution is usually to try to add the new "fact" to the symbol system's "knowledge" -- but that can keep going on forever. And remember (lest you think we ourselves have frame problems too) that when the system falls into the frame problem, it is not just slightly wrong, it is radically wrong, so wrong that all prior bets are off and it's not clear that it ever really "knew" anything at all.
Could the frame problem be a variant of the credit/blame
assignment problem, but a radical one because the system is not learning,
online, but meant to already "have" the knowledge in its head?
So could the frame problem be another symptom of the problem of trying to do AI with symbols only? Because of course learning is mostly sensorimotor and nonsymbolic.
How is it different in our heads? How come the symbols in our minds mean something? Perhaps it's because some of them (not all, just some) are not just defined in terms of still further symbols (that would go on forever without reaching a meaning), but because they are instead connected to the things they stand for by the sensorimotor mechanisms that detect and recognize those things?
In other words, symbolic know-that needs to be grounded in sensorimotor know-how. And much of that sensorimotor know-how comes from sensorimotor learning. (After that, once you have grounded a basic vocabulary, the rest could all be gotten by combining and recombining the symbols into higher-order categories, the way dictionary definitions too -- but probably even there, all those abstractions need to be "refreshed" by some direct sensorimotor connections now and then.)
So perhaps the frame problem is a symptom of the symbol grounding problem; and perhaps that's why Searle's argument works too.
Once a system is grounded, it inherits all the power of the Turing Machine and computation, and the rest can be done with just symbols and algorithms.