Copyright 1999 by Donald R. Tveter, http://www.dontveter.com, commercial use is prohibited. This material cannot be quoted at length or posted elsewhere on the net or included in CD ROM collections. Short quotations are permitted provided proper attribution is given.
Most people interested in AI BELIEVE that a digital computer can do all that people can do if it is programmed correctly. Meaning that there are plenty of arguments over what the correct method of programming will turn out to be. Also meaning that some people doubt digital computers can do the job.
Symbol processing is the original dominant BELIEF of the AI community. Symbol processing requires the use of symbols, structures of symbols and rules.
Symbols are defined as unique marks as on a piece of paper. Symbols can be equal or not equal but there are no other relations defined between them. Strictly speaking then, numbers are not allowed because numbers have an ordering to them, one unique number will be greater than or less than some other number. In practice symbol processing programs often incorporate the use of numbers so most "symbol processing" systems go beyond the processing of symbols. Implicitly they are saying that the statement, "symbol processing is necessary and sufficient for general intelligent action" is clearly false. They should go around saying that "symbol processing and numerical computing are necessary and sufficient for general intelligent action". But symbol processing believers go around saying "symbol processing is necessary and sufficient for general intelligent action" anyway.
The second part of the doctrine of symbol processing is that you need structures of symbols, you need to be able to glue them together into more complex structures as well as tear the structures apart.
The third part of the doctrine of symbol processing is that you need rules.
(A critique of symbol processing is in order and I'm working on one. Anybody know of any other worthwhile critiques, especially online ones?)
Stuart and Hubert Dreyfus are well-known critics of AI, they have an article online aimed particularly at criticizing the use of rules in symbolic expert systems.
Symbol processing has not been doing very well with respect to producing systems that can deal with the whole real world. Some people doubt that it ever will thus prompting some researchers to look for alternatives. The current popular alternative BELIEF is "neural networking". Here you have a very large number of simple processors all working together to compute some result. In many cases you get some nice human-like behavior out of the systems especially the ability to learn, something that symbol processing is poor at. The emphasis in this approach is the use of real numbers, indeed as far as I can tell the basic concept of neural networking applied to AI is that "real numbers are necessary" because while the architecture of a neural computer is different than that of a digital computer all the calculations you can do with the network can be done by digital computers ASSUMING your neural network uses limited precision numbers (using analog circuits is another issue, see Computability below). For a while believers in neural networking were saying that "AI (meaning symbol processing AI) is dead". Now it seems that both sides have mellowed a bit and many are willing to accept whatever works including hybrid systems that use symbol processing and neural networking concepts. Marvin Minsky has an online paper dealing with his belief that AI needs to combine symbol processing with neural networking.
One branch of computer science is interested in what can possibly be computed by a computer or the brain. The dominant BELIEF in AI is that the brain is effectively the equivalent of a digital computer and therefore digital computers can do what brains do, its only a matter of figuring out how and producing a fast enough piece of hardware to run the algorithms. But there are dissidents here as well, that come from two camps, the first group is the neural networking community and the second group consists of physicists who advocate that quantum computing is necessary to reproduce human abilities.
The nicest analysis of the difference between symbol processing and neural networking that I know of comes from computer scientist Bruce MacLennan. MacLennan argues that neural analog computing has some properties that digital computing does not have. He has some very readable articles available on his position, see [MacLennan 1991a] and [MacLennan 1991b]. A less readable article (due to the use of math) is [MacLennan 1993]. This last article briefly mentions how some analog systems can do things that Turing machines (and therefore digital computers) cannot do.
Another analysis of the capabilities of analog computers has been done by Hava Siegelmann, for a short summary see the article in EE Times: Analog computer trumps Turing model, for more details see Hava Siegelmann's home page. At this point in time the experts don't think its possible to ever implement analog computers with their super-Turing capability. But this only makes me think of one of Arthur C. Clarke's famous comments that went something like: "If a distinguished but elderly scientist says something can be done he is probably right on the other hand if he says somthing cannot be done he is almost certainly wrong".
Roger Penrose, a theoretical physicist has written a couple of books: The Emperor's New Mind and Shadows of the Mind where he argues that people are capable of insights into the truth of a proposition without being able to prove it algorithmically. Some people see that Penrose is obviously right. Some people see that Penrose is obviously wrong. What's obvious then is that the issue is not obvious. Unless you're a specialist in this area I recommend avoiding it.
But besides the computability argument Penrose is among a number of researchers proposing that there is quantum computing going on in the brain and quantum effects are responsible the flash of insight phenomenon. Penrose proposes that quantum computing is happening in the microtubules of cells. Microtubules are hollow tubes that serve as a skeleton for a cell (cytoskeleton). They can take on two different states called alpha and beta tubulin depending on the presence or absence of an electron potentially making for a nice digital system but one that Penrose and others propose is controlled by quantum effects. I think this idea should be taken more seriously than the theoretical argument about computability since the properties microtubules can actually be researched whereas arguments about computability can only be argued. More on quantum computing below.
The traditional BELIEF in AI and in the scientific community in general is that all the important action in the brain takes place using neurons that function as switches. The usual estimate is that there are 10^11 neurons in the brain with an average of 10^3 connections to other neurons with each neuron firing at up to 100 times per second giving what I will call a bit twiddling rate of 10^16 bits per second. TRADITION says that this is sufficient to account for human thinking and since digital computers are fast coming up to this rate of bit twiddling it won't be long before we have a digital computer with the power of the brain.
With the only other alternative to explain thinking being that a human soul or spirit is involved the neuron as a switch theory did not have any competition and so it became THE EXPLANATION. But a new candidate has appeared on the scene: the cytoskeleton. It is a collection of hollow fibers (microtubules) made out of a protein called tubulin. Until recently the only known use for the cytoskeleton was as a skeleton that maintains the shape of a cell and in some cases the tubes extend beyond the cell and are the cilia used for instance in paramecia to enable them to swim around. The microtubules consist of molecules of tubulin that can be in two different states depending on the presence or absence of an electron, a nice digital system.
Studies of paramecia seem to show that they can learn:
For example, a number of studies have observed paramecia swimming and escaping from capillary tubes in which they could turn around. In general, results showed that with practice the ciliates took successively less and less time to escape, indicative of a learning mechanism ( French, 1940; Applewhite and Gardner, 1973; Fukui and Asai, 1976). Many other experiments suggest paramecia can learn to swim in patterns and through mazes and have a short-term memory, although some of these behaviors depend on their environment ( Applewhite, 1979;). [Hameroff et. al. 1993]
If neurons are responsible for learning in multi-celled animals it is hard to explain how a one-celled animal with NO neurons can learn. The theory is that the cytoskeleton is the nervous system of the paramecium and the cytoskeleton is a miniature computer. The gist of this particular article is to explore ways of doing computing using microtubules without considering quantum mechanical effects. The authors of that article also estimate that a paramecium (or a neuron or some other cell) could twiddle bits at the rate of 10^13 bits per second. Thus every cell with microtubules (this is almost all cells) may contain a computer. A human brain counting only the use of 10^11 neurons and allowing for some redundancy would twiddle 10^23 bits per second. Another estimate I've seen is 10^28 bits per second. In either case this is rather a lot more than a digital computer can manage at this time and if this is what is going on then it will be a while before digital computers can compete with the brain.
Needless to say people with the traditional BELIEF that computation in the brain is done with the neuron as a switch model don't believe the above scenario even one little bit.
If the brain is a quantum computer then it could impact AI in two possible ways. First, there may be something that a quantum computer can do that a digital computer cannot do thereby making human level AI via digital computer impossible even in principle. Second, maybe a digital computer could in principle do whatever a quantum computer can do except given that quantum computation is so much faster than digital computation that in practice the digital computer would never be able to compete. In either case the grand goal of making a computer roughly equal to a human being would have to await the construction of quantum computers as complex as the brain.
Consciousness is an unexplained phenomenon. People have it and it is generally believed that higher level animals have it. But what is it? How would you know when you see it? Some people speculate that we will only have true artificial intelligence if the system is conscious, meaning that digital computers will never be able to do all that conscious people can do. Various philosophers and scientists write books and papers on the subject that I've ignored due to my bias toward a quantum mechanical interpretation of the mind/brain so as of right now I have little to say about the non-quantum philosophies of consciousness. (HELP!).
Philosopher David Chalmers has a collection of Online papers on consciousness
The oddest theory of consciousness I know of comes from computer scientists who BELIEVE (or say they believe) that a digital computer executing a program is automatically conscious. (I really would like to see documented cases of people claiming this.)
The idea that quantum mechanical waves and fields could be responsible for consciousness goes back to the 1920s and was suggested by at least physicist James Jeans and biologist Alfred Lotka. There are a number of different quantum mechanical approaches to consciousness, too many to try and explain here.
The proposal for quantum consciousness that gets the most attention comes from physicist Roger Penrose and physician Stuart Hameroff and it involves the use of microtubules in the brain to support consciousness. Penrose proposes that consciousness comes from an orchestrated collapse of the wave function. For more on their theory see the books: The Emperor's New Mind and Shadows of the Mind Also see Stuart Hameroff's Web Page.
A particularly nice online paper by physicist Dimitri Nanopoulos is [Nanopoulos 1995]. It considers quantum mechanics, string theory, microtubules and brain function and brings in the psychology of William James and Sigmund Freud in a 73 page paper. Nanopoulos may already have found what Penrose and Hameroff are looking for. It may be right, it may be wrong but the scope of this paper is so stunning you don't want to miss it.
Physicist Jack Sarfatti is a frequent contributor to comp.ai.philosophy with his theory of consciousness that requires that standard quantum mechanics must be altered slightly to make consciousness possible. Jack's preferred perspective on QM uses the interpretation of physicist David Bohm as the starting point. Bohm split the Schroedinger wave equation into two parts, one that covers matter in a classical way and a second part that is a quantum force (also referred to as a pilot wave) that controls the path of a particle. This interpretation of quantum mechanics is a no-collapse theory so it differs from the Penrose proposal. Jack's proposal is that particles can generate a back-action on the quantum force and this is responsible for consciousness. More by Jack can be found on his sites:
Another online source is the Archives of QUANTUM-MIND@LISTSERV.ARIZONA.EDU
I've attempted to maintain web pages on the subject of Quantum Mechanics and Natural Intelligence, unfortunately at this time it is in serious need of being updated.
Descartes is usually given the credit for introducing the idea that there are two kinds of things in the world, matter and "mind stuff" (spirit). This is a belief called dualism. Also before Descartes theologically oriented individuals have typically claimed that spirit is "supernatural" and cannot be studied by science. Scientists quite often dismiss the idea that there is any "supernatural" spirit and eagerly ridicule people (even other scientists) who believe it or even propose it. Except of course the "supernatural" may not be "super" at all, it may just be the result of an extremely advanced science and technology. (Recall Arthur C. Clarke's famous comment that "Any sufficiently advanced technology is indistinguishable from magic".) One idea is that quantum mechanical waves and fields may be the spiritual portion of the universe and so science may already be studying spirit without really knowing it.
"Learning in Protozoa" by PB Applewhite in Biochemistry and Physiology of Protozoa Volume 1 edited by M Levandowsky and SH Hunter, Academic Press, 341-355, 1979.
"Tube Escape Behavior of Paramecia", by PB Applewhite and FT Gardner, in Behav Biol 9:245-250, 1973.
"Trial and Error Learning in Paramecium" by JW French in J Exp Psychol 26:609-613, 1940.
"Spiral Motion of Paramecium Caudatum in a small capillary Glass Tube" by K Fukui and H Asai in J Protozool 23:559-563, 1976.
Hameroff 1993, Stuart Hameroff, Judith E. Dayhoff, Rafael Lahoz-Beltra, Steen Rasmussen, Ezio M. Insinna and Djuro Koruga, "Nanoneurology and the Cytoskeleton: Quantum Signaling and Protein Conformational Dynamics as Cognitive Substrate", in Rethinking Neural Networks: Quantum Fields and Biological Data, editor Karl H. Pribram, International Neural Network Society Press and Lawrence Erlbaum Associates, 1993.
MacLennan 1991a, Bruce MacLennan, "Continuous Symbol Systems: The Logic of Connectionism", from the Ohio State neuroprose archive
MacLennan 1991b, Bruce MacLennan, "Characteristics of Connectionist Knowledge Representation", from the Ohio State neuroprose archive or from netlib.org wherever that is.
MacLennan 1993, Bruce MacLennan and David Wolpert, "A Computationally Universal Field Computer That is Purely Linear", from the Ohio State neuroprose archive.
Nanopoulos 1995, Dimitri Nanopoulos, "Theory of Brain Function, Quantum Mechanics and Superstrings" by Dimitri Nanopoulos available from the Los Alamos National Laboratory Physics preprint archives.
Penrose 1989, Roger Penrose, The Emperor's New Mind, Oxford University Press, 1989.
Penrose 1994, Roger Penrose, Shadows of the Mind, Oxford University Press, 1994.