Date: Sun 13 Nov 1988 21:51-EST From: AIList Moderator Nick Papadakis Reply-To: AIList@AI.AI.MIT.EDU Us-Mail: MIT LCS, 545 Tech Square, Rm# NE43-504, Cambridge MA 02139 Phone: (617) 253-6524 Subject: AIList Digest V8 #125 To: AIList@AI.AI.MIT.EDU Status: R AIList Digest Monday, 14 Nov 1988 Volume 8 : Issue 125 Philosophy: Epistemology of common sense The study of intelligence Computer science as a subset of artificial intelligence Lightbulbs and Related Thoughts IJCAI Panels ---------------------------------------------------------------------- Date: Mon, 7 Nov 88 11:21:08 EST From: "Bruce E. Nevin" Subject: epistemology of common sense In AIList Digest for Monday, 7 Nov 1988 (Volume 8, Issue 121), in a message dated 31 Oct 88 2154 PST on the topic "AI as CS and the scientific epistemology of the common sense world", John McCarthy has persuasive words for colleagues who prefer to limit their research to things that are amenable to tidy mathematical formulation. The audience of "neats" he was addressing should ignore this. I want to talk about aspects of common sense that seem even less tidy. (But there is hope, cf. references at the end.) JMC> Intelligence can be studied | . . . | (3) through studying the tasks presented in the achievement of | goals in the common sense world. | . . . | I have left out sociology, because I think its | contribution will be peripheral. | . . . | AI is the third approach. It proceeds mainly in computer science There is more to common sense than the study of tasks and goals specified in physical terms. Much of common sense involves social facts, not just physical facts. A telltale of social facts is that they are matters of convention. Absent intelligent agents conforming to them, they do not exist. Restricted to physical facts, common sense concerns things like "I can't put the blue pyramid in the box, it's already in there" or "I can't put the lintel on yet, I need to move the second column closer to the first." Suppose we had an AI equipped with common sense defined solely in terms of physical facts. This is somewhat like the proverbial person who knows the price of everything but the value of nothing. We deceive ourselves when we put labels on things like "road" or "vehicle" or even "arch" in a knowledge base. We have many expectations and other associations with these terms that a knowledge base lacks--unless we explicitly include those associations. If and when we do begin to include such associations (that line defines my lane, this is the slow-speed lane, drive on the right--unless in England or Sweden or . . . that joker's trying to pass me in the breakdown lane . . . this must be Boston . . . ) we are involved with the sociology of knowledge. Look at Erving Goffman on, say, presentation of self or interaction rituals. Look at W. Pearce (UMass Amherst) on communication rules and rules for constituting the social order. For starters. An AI must be responsive as a member of the social order if it is to be regarded as intelligent by humans. It does not need the physiological or psychological mechanisms of humans, but it does need to understand their conventions. Bruce Nevin bn@bbn.com ------------------------------ Date: Mon, 7 Nov 88 10:22:14 PST From: norman%ics@ucsd.edu (Donald A Norman-UCSD Cog Sci Dept) Reply-to: danorman@ucsd.edu Subject: The study of intelligence Time for comment from a Cognitive Scientist on the appropriate approach to the study of Intelligence. As usual, John McCarthy has provided us with a cogent and coherent analysis of the approaches one might take, but although his approach appears sensible, I wish to disagree about the importance of several aspects he downplayed. McCarthy states: Intelligence can be studied (1) through the physiology of the brain, (2) through psychology, (3) through studying the tasks presented in the achievement of goals in the common sense world. True enough, except that I would add several others: (4) through an analysis of intelligent behavior (in the abstract, as is most frequently done in philosophy, and in some AI and Cognitive Science endeavors) (5) Through an analysis of how intelligent behavior results from an interaction of individual cognition, the cognitions of others, the social structures and cultures, and the physical environment, [In part, what we here at UCSD call "Distributed Cognition," which is highly related to the recent work on "Situated Action" (See Lucy Suchman's book or the papers of Agre and Chapman, for example).] Real intelligence takes place as an interaction among people, in a social environment, constrained by the particular experiences of the participants and by the biological structures of the organism (not just the brain, but also the sensory systems, the locomotive and grasping mechanisms, and the whole regulatory system which interacts dramatically with our cognitions. Traditional analyses of intelligent behavior leave out the role of emotions, of limited sensory and reasoning capabilities, of the example-driven aspects of interpretation and memory retrieval and decision making. These analyses make logical sense and can lead to the development of intelligent machines, but they are not accurate portrayals of human intelligence. They also (and as a direct result) miss the creative aspect of human intelligence and fail to characterize properly real human behavior, both the insightful variety, and the class of things called "human error." McCarthy talks of "common sense" but has he really studied what common sense is about? One person's common sense is another's nonsense. Common sense varies widely from culture to culture. I highly recommend the paper by Geertz (an anthropologist -- one field McCarthy left out): Geertz, G. (1983). Local knowledge: Further essays in interpretive Anthropology. New York: Basic Books. (Especially see the essay "Common sense as a cultural system," pp. 73-93.) In conclusion: John McCarthy has given a logical set of procedures to follow in the study of Artificial Intelligence. They make sense and will lead to advancement in the understanding of one form of Artificial Intelligence. But there are many possible forms of Artificial Intelligence, and it is highly likely that dramtically different other approaches will also prove fruitful. However, I am interested in Real Intelligence, and for this domain, McCarthy's approach is much too limited, for it neglects the powerful and important contribution of biological structure, of social interaction, of the role of cultural knowledge, and of the interaction among individuals and the environment. We work in a world of incomplete and erroneous knowledge, ambiguous situations and communications, and partial specifications of all sorts, where much of behavior is driven by the accidents of the environment or by biological needs and limits. And almost all of our intelligent behavior results from social interaction and by the use of artificial artifacts (which, of course, were created by us to aid our thought and communication proceses -- cognitive artifacts, I call them). We can only study Real Intelligence by studying Real Organisms in interaction with other organisms, their cultural knowledge, and their environment. don norman Donald A. Norman [ danorman@ucsd.edu BITNET: danorman@ucsd ] Department of Cognitive Science C-015 University of California, San Diego La Jolla, California 92093 USA UNIX: {gatech,rutgers,ucbvax,uunet}!ucsd!danorman [e-mail paths often fail: please give postal address and all e-mail addresses.] ------------------------------ Date: 8 Nov 88 01:46:40 GMT From: quintus!ok@Sun.COM (Richard A. O'Keefe) Reply-to: quintus!ok@Sun.COM (Richard A. O'Keefe) Subject: Re: Computer science as a subset of artificial intelligence In a previous article, Ray Allis writes: >I was disagreeing with that too-limited definition of AI. *Computer >science* is about applications of computers, *AI* is about the creation >of intelligent artifacts. I don't believe digital computers, or rather >physical symbol systems, can be intelligent. It's more than difficult, >it's not possible. There being no other game in town, this implies that AI is impossible. Let's face it, connectionist nets are rule-governed systems; anything a connectionist net can do a collection of binary gates can do and vice versa. (Real neurons &c may be another story, or may not.) ------------------------------ Date: 10 Nov 88 12:23:41 GMT From: oodis01!uplherc!sp7040!obie!wsccs!dharvey@tis.llnl.gov (David Harvey) Subject: Re: Lightbulbs and Related Thoughts In a previous article, Tony Stuart writes: > > On a similar track, I have often thought that once we find a > solution to a problem it is much more difficult to search for > another solution. Over evolutionary history it is likely that > life was sufficiently primitive that a single good solution was > sufficient. The brain might be optimized such that the first > good solution satisifies the problem seeking mode and to go > beyond that solution requires concious effort. This is an > argument for not resorting to a textbook as the first line of > problem solving. > Usually, advances by humans comes on top of what has gone before, not inside a vacuumn. I realize that this is not exactly what you intended to present here, but it comes out that way regardless. As to the better solution, that usually is the way it happens. For examples, consider Keppler seeing inconsistencies between the model proposed by Aristotle and the calculations (just think how much faster his work would have been with a computer!) he made. This of course prompted him to devise a new model. Galileo and Newton also saw inconsistencies between what was commonly believed and the effects of gravity, ie, that accelaration was a constant not affected by the mass of the object. Einstein saw inconsistencies even in this model and developed the theory of relativity. In other words, these people KNEW the textbook solutions. What characterized them as being different from the masses is that they had the tenacity to reject the 'textbook' solution when a better model came to mind. Just how this can be emulated in a computer is not that easy. The only thing that can be said is that inconsistencies of data with the rule base must allow for a retraction of the rule and assertion for new ones. > > I've often wondered about the differences between short term > and long term memory. > Don't forget to include the iconic memory. This is the buffers so to speak of our sensory processes. I am sure that you have saw many aspects of this phenomenon by now. Examples are staring at a flag of the United States for 30 seconds, then observing the complementary colors of the flag if you then look at a blank wall (usually works best if the wall is dark). There are other ways of observing that there really is such a thing as iconic memory, but these must be performed in a lab setting with blind studies. I helped perform one of these at the University of Utah. How do you implement this into your model? I don't know, and I doubt anyone else does either since much research must yet be done to see the relationship between iconic, short term, and long term memory. Also, the differences between conscious and subconscious memory processes must be considered. Much of this iconic information makes its way into memory via the subconscious track, which I would cite as evidence the studies being performed by various researchers in Psychology. You have observed the linking process that takes place in our long term memories. This is of course a dandy model until you begin to look at some of our links. They have some of the following characteristics: [1] Some of them seem to link together totally randomly. I am sure you have observed the phenomenon that some of your own links are rather mysterious, where the items are not logically related at all. Nevertheless most of them ARE logically related. Maybe we can randomly throw in a time frame for the other links. This of course supposes that we can prove that time is indeed the model that determines them. By time I mean close time proximity for the linked structures. [2] There are a massive amount of them that we search, sometimes in vain. As witness to this consider the tip-of-the-tongue phenomenon that we are cursed with. I am sure that we all have experienced it. Perhaps those with photographic memory are not cursed with it, but not being so blessed I would not know. Also, some of these sturctures unlink with time and fall away. This last tidbit of course goes against the conventional textbook wisdom that they stay there forever. [3] Since there are so many, we MUST use parallel processing to search them all. Also realize that they are massive in nature, perhaps to the point of exceeding most mass storage devices (disks) in use today. The short term memory does not necessarily have to have a different data representation. It still has a linking type nature. The main difference I see between the two is that short term memory has far fewer links than long term. What needs to be done is to study why and how this short term memory links up with the long term memory. Perhaps frequency of use could be researched as the causal factor. Initially, we must establish a linking base for these short term facts to attach themselves to. As I see it there are several ways it will link into the established long term memory. First of course is the logical link. Another would be a time frame link where what was considered immediately before or after would be what we attach it to. Also, since it is a well established fact that we can chain things much better via poetry than prose, rhythm and actual morphemes must be considered for chaining. > A side effect of this model is that information in short > term memory cannot be used unless there is a hole in the long > term memory. This leads to problems in bootstrapping the > process, but assuming there is a solution to that problem, it > also models behavior that is present in humans. This is the > case of feeling that one hears a word or phrase a lot after > he knows what it means. Another part of the side effect is > that one cannot use information that he has unless it fits. > This means that it must be discarded until the long term > memory is sufficiently developed to accept it. > The problem is that there are more than enough holes for something to be fit in. Inconsistency seems to thrive in human beings. It is only when new information conflicts enough with old that we attempt to rationalize the two conflicting 'facts'. Unless the new information outweighs the old in some way it never replaces it. It can and does coexist with the old in tension in many cases. It is only when we reach the discomfort level that we attempt to resolve the disparity of the two in our fact base. Well, now that I have given enough for Psychologists and AI researchers to work on for the next 50 years (:-) I can go back to such mundane chores as homework and sleeping. Hmm, are we going to model the activity of sleeping in our machine? dharvey@wscss The only thing you can know for sure, is that you can't know anything for sure. ------------------------------ Date: 10 Nov 88 18:36:41 GMT From: umix!umich!itivax!ttf@uunet.UU.NET (Fejel) Subject: IJCAI Panels At the IJCAI Local Arrangements Committee meeting this past Friday, we were urged to submit panel suggestions. I have noticed that the net explodes whenever AI "infringes" on people's values. It seems that ethical and moral issues generally stir up much interest and controversy, though unfortunately it is often the case that more heat than light is generated. With this in mind, I would like to propose a panel discussion on AI, Ethics, and Morality. It could have three viewpoints: 1)The ethics of AI application (ie. defense-related domains). 2) The ethics of building artificial persons (as presented by Michael LaChat)(admittedly a bit blue-sky), and most important, 3) The view of ethics and morality as cognitive tasks, and therefore legitimate objects of research within the AI community. If you're interested, and especially if you are thinking of coming to Detroit, please email me back your comments and suggestions, and maybe send a panel discussion request to the IJCAI organizing commmittee. Tihamer P.S. Unfortunately, I am not familiar with any work being done in this domain. Anyone have any pointers? arpanet: ttf%iti@umix.cc.umich.edu uucp: ...{pur-ee,well}!itivax!ttf (Tihamer T. Toth-Fejel) Industrial Technologies Institute, Ann Arbor, Michigan 48106 work phone: (313) 769-4248 or 4345 *----*----*----*----*----*----*----*----*----*----*----*----*----* ------------------------------ End of AIList Digest ********************