From ashwin@gatech.edu (Ashwin Ram) Subject: Class notes - CAN MACHINES THINK Date: Sun, 6 Jun 1993 20:17:36 GMT CONCLUSIONS(?): %---------------------------------------------------------------------------- Basic question: Can computers be intelligent? Can computers think? Note: This is a different question from ``Can machines think?'' ``Are humans machines?'' ``Can machines think like humans?'' ``How would we know whether a computer is thinking? Keep in mind that it is not demeaning to humans to say that computers can think or that humans are (physical/biological) machines. Humans are certainly (very complex) physical/biological devices, but that does not demean us in any way. If it should turn out that our intelligence was indeed the result of a (very complex) program, that would not demean us either. %---------------------------------------------------------------------------- Related question: What is intelligence? What is thinking? 1. A biochemical activity. (Does a neuron think?) 2. A neurobiological activity activity. (How many neurons are necessary?) 3. A non-computational "physical" activity. (What is it?) 4. A non-physical "mental" activity. (How could we ever know this?) 5. A behavioral activity. (Is I/O behavior sufficient?) 6. A "functional" activity. (The AI view: goals, plans, reasons, hypotheses, explanations, memory. Self-reference? Judgement? Emotion?) %---------------------------------------------------------------------------- Can computers think? Some common answers: 1. No (dualist/mystic): Computers lack "mental stuff". They don't have intuitions, feelings, phenomenology. (Soul?) 2. No (neurophysiology critical): Even if their behavior is arbitrary close, biology is essential. 3. No (beyond our capabilities): Not impossible in practice, but too complex. There are limits on our self-knowledge which will prevent us from creating a thinking computer. 4. Yes (but not in our lifetimes): Too complex. Practical obstacles may be insurmountable. But with better science and technology, maybe. 5. Yes (functionalist): Programs/computers will become smarter without a clear limit. For all practical purposes, they will 'think' because they will perform the FUNCTIONS of thinking. (The standard AI answer.) 6. Yes (extreme functionalist; back to mystic): Computers already think. All matter exhibits mind in different aspects and degrees. 7. Others? %---------------------------------------------------------------------------- Clocks vs. Computers vs. People (Winograd and Flores) CLOCKS COMPUTERS 1. Apparent autonomy: little lot 2. Complexity of purpose (not design): simple several purposes 3. Structural plasticity: constant changing 4. Unpredictability: little lot Computers are tending more and more towards "minds". %---------------------------------------------------------------------------- Alan Turing anticipated the following objections (as discussed by Hofstadter): 1. Theological objection: Thinking is a function of man's immortal soul. 2. Heads In The Sand objection: The consequences of machines thinking would be too dreadful. 3. Mathematical objection (Lucas's argument): [See below.] 4. Argument from Consciousness: Machines will never be self-conscious. 5. Disability argument: It is impossible to make a machine do X, where X = be friendly, fall in love, have a sense of humor, do something really new, ... 6. Lady Lovelace's objection: "[Babbage's] Analytical Engine has no pretensions to _originate_ anything. It can do _whatever we know how to order it_ to perform." 7. Informality of behavior: Since there are no rules of conduct running men's lives, men aren't machines. 8. Extra-Sensory Perception argument: You could tell a machine from a man with ESP, because a machine could at best guess randomly. %---------------------------------------------------------------------------- Some Serious Arguments Against AI: Mathematical arguments Biological arguments Simulation arguments Homunculus arguments Creativity/predictability arguments Artifact arguments Symbol grounding arguments %---------------------------------------------------------------------------- Mathematical arguments (The Lucas Argument, as discussed by Hofstadter in Goedel, Escher and Bach; also discussed in the The Abilities of Men and Machines paper by Dennett): All machines are Turing Machines (universal computers), but humans are not; they are more than Turing Machines. Because: Suppose Jones (a particular human) was a realization of a Turing Machine T. Then by Goedel's theorem, there is something A that he can't do (namely, prove T's Goedel sentence). But since Jones can do A, he can't be a realization of a Turing Machine. Refutations: 1. Dennett: Provable limitations aren't limitations on what can be done heuristically with high reliability. Analogy: All computer programs are algorithms. (true) There is no feasible algorithm for checkmate in chess. (true) Therefore checkmate by computer is impossible. (false) 2. Hofstadter: There are things humans can't do too. Can we really Goedelize? We know it can be done, but can Jones do it? 3. Hofstadter: Goedel's argument applies to lower levels of AI programs, which are simple, formal, logical systems. But higher levels are models of the mind, and can be informal just like the mind is. This is where the intelligence lies. So human-like intelligence could emerge out of formal lower levels. Machines with multiple levels of knowledge, including "strange loops" in which a level of description applies to itself, are fundamentally different to what we normally think of as "machines" (which have no meta-knowledge). %---------------------------------------------------------------------------- Simulation arguments: A simulation is not the real thing. You wouldn't expect to get wet in the presense of a simulation of a hurricane. Similarly, although a simulation of intelligence would tell you a lot, it would not really _be_ intelligence. Refutations: 1. Hurricanes and intelligence are fundamentally different. A simulation of intelligence would in fact display intelligence. 2. Dennett, in the Why You Can't Make A Computer That Feels Pain paper: Does a simulation need to be indistinguishable from the real thing? The hurricane simulation would give you good descriptions of hurricanes, and predictions about a hurricane's behavior. So also in AI we seek a theory of intelligence. We are not trying to prove that humans are computers (any more than hurricanes are), but rather trying for a rigorous theory of human psychology. The program _instantiates_ the theory. %---------------------------------------------------------------------------- Homunculus arguments (Hume's problem, as discussed by Dennett in the AI As Philosophy And As Psychology paper): We can't account for perception unless we assume it gives us an internal image or model of the world. But what use is this image unless we have an inner eye to perceive it? And then how do we explain its capacity for perception? Analogous homunculus problem in AI: The only psychology that could explain intelligence must posit internal representations -- ideas, sensations, impressions, maps, schemas, propositions, neural signals, whatever (radical behaviorism excluded). But nothing is intrinsically a representation of anything; it is a representation _for_ or _to_ someone. Any system of representations must have a user or interpreter external to it. This interpreter must have psychological abilibities; it must understand the representations; it must have beliefs and goals. But this is a homunculus. Therefore psychological without homunculi is impossible. Refutation: AI has solved this problem through reductionism. We reduce intelligence to smaller and stupider "homunculi", until ultimately they begin to look less like homunculi and more like machines. In other words, the internal "modules" of intelligence need not be full blown intelligences in themselves, so they can ultimately be reduced to machines. %---------------------------------------------------------------------------- Creativity and predictability arguments: Computers can only do what they are programmed to do. They are not intelligent; they are merely following the instructions of a program. Refutations: 1. Evans: The same is true of animals and men; they are "structure determined". Humans may well be the same way. If you could model humans down to the level of physics (the program will probably need more memory than there are atoms in the universe, but ignore that for a moment), they might be predictable too since ultimately they are following the laws of physics. But that apart, if we could understand the human "program", we might make the same objection to human intelligence. Computers are predictable only in theory (just like humans); once we succeed in building an intelligent computer, it will be large and complex enough that it will for all practical purposes be as unpredictable and as non-instruction-following as humans are. Note: it is not important to this argument who wrote the program, whether it evolved through biology, or whether it was written by humans and/or other computers. 2. Computers that learn from their experiences will not only be doing what they are programmed to do; they will be able to go beyond their initial programs. And since their experiences will not be deterministically known in advance (and, for any given machine, you may not know its past experiences), they will not be predictable. %---------------------------------------------------------------------------- Artifact arguments: Computers are artifacts; we created them, so they can't be intelligent. Refutation: We create babies too, yet we are willing to grant intelligence to babies. %---------------------------------------------------------------------------- Searle's Chinese Room: Refutations: 1. McCarthy: "Searle confuses the mental qualities of one computational process, himself for example, with those of another process that the first process might be interpreting, a process that understands Chinese, for example." It is the program that the computer is executing that understands Chinese. Roughly, your neurons or your brain don't understand English, but rather the program they are executing (your "mind") does. Similarly, Searle doesn't understand Chinese, but the Chinese Room does by virtue of its software. McDermott: There are two understanders here, Searle and Searle simulating the Chinese understander. It is the latter that understands Chinese. 2. Hayes: The basic flaw in Searle's argument is a widely accepted misunderstanding about the nature of computers and computation: the idea that a computer is a mechanical slave that obeys orders. This popular metaphor suggests a major division between physical, causal hardware which acts, and formal symbolic software, which gets read. This distinction runs through much computing terminology, but one of the main conceptual insights of computer science is that it is of little real scientific importance. Computers running programs just aren't like the Chinese room. Software is a series of patterns which, when placed in the proper places inside the machine, cause it to become a causally different device. Computer hardware is by itself an incomplete specification of a machine, which is completed - i.e. caused to quickly reshape its electronic functionality - by having electrical patterns moved within it. The hardware and the patterns together become a mechanism which behaves in the way specified by the program. This is not at all like the relationship between a reader obeying some instructions or following some rules. Unless, that is, he has somehow absorbed these instructions so completely that they have become part of him, become one of his skills. The man in Searle's room who has done this to his program now understands Chinese. 3. If the computer has _real_ experiences (it interacts with the world), if it can connect the Chinese characters it is manipulating with these experiences, if its symbols are grounded in the real world, then it _is_ understanding Chinese. Searle's Chinese Room needs sensory and motor devices. %---------------------------------------------------------------------------- Harnad's Symbol Grounding Problem: Resolutions: 1. Powers: These results suggest three possible resolutions of the symbol grounding problem: the symbol/non-symbol distinction is not meaningful; neural networks can exhibit 'symbolic' behaviour and structure; and, a sensory-motor environment can provide grounding. %---------------------------------------------------------------------------- Finally, a moral/ethical issue: Let suppose that it is possible to build intelligent computers. Should we? Would we leave critical decisions about (human) life, liberty, and happiness to intelligent computers? Would intelligent computers be used by some for ``evil'' purposes? Would intelligent computers take over the world? What, if any, is the responsibility of AI researchers? [Some in AI have claimed that their loyalty is to intelligence, not to humanity.] What should WE as students of AI DO? %---------------------------------------------------------------------------- Copyright (c) Ashwin Ram, 1990-93 Assistant Professor, College of Computing Georgia Institute of Technology, Atlanta, Georgia 30332-0280 E-mail: