2pat S ZsXr/tiVO/O - Stacks are the Stanford
Transcription
2pat S ZsXr/tiVO/O - Stacks are the Stanford
, <BARR>MK2. 1 f*i/OocO/£Zx?2— Tue 6-Feb-79 9: 4IPM /? 2pat S ZsXr/tiVO/O KNOWLEDGE . crown ( 10, 10, 10) After discussing what meta-knowledge might mean, and what kinds of things can be "known about knowledge", we review some MX is psycho studies for a feeling of the epistomology of MX. an intrinsic part of human cognition Review the reasons for developing techniques for rep. of MX in All as well as extent/limits of those techniques. . endcrown OMeta-knowledge is knowledge about knowledge. > begin to explore exactly what this sentence might mean. The common understanding is something like this This paper will . crown (10, 10, 10) Knowledge is an accumulation of some sort, of "facts" about the world. Meta-knowledge is the same kind of accumulation of but instead of being about the world, they are about other other knowledge. . endcrown .skip cont inue This explanation appeals to our intuitive notion of knowledge being <+about the world>. referring to a morphism between the world of our perceptions and our storage of information. It is my belief that these intuitions are no longer useful in the study of cognition, that in fact they lead to serious confusion. In going through examples of Al systems and psychological studies that explore meta-knowledge. the difficulties will be examined. - see note on why MX breaks the circle. . it to be "about". . there is nothing for The use of meta-knowledge in expert Al systems like MYCIN for facilitation of explanation and acquisition is, a key We will discuss breakthrough in the design of intelligent systems. why later, but first let's deal with an easy preliminary issue. . glitch You have to discuss the point that "knowing not", for instance, best described as a "fact" in the head. But then again, neither is not is "snow is white". endg l i tch . <BARR>MK2. I Tue 6~Feb-79 9: 4IPM What is Knowledge? "Knowing" is intimately tied to behavior, and thus to process. When we say that an animal knows something, we mean that it acts in a certain way that is different than it would if it didn't. The fact that there is some piece of knowledge or fact or is inferred by us: Used to explain the difference (change in the animal) in a simple way that we all understand. Pseudo-process (Giraffes) In Al, we have established for ourselves that the meaning of a representation is intimately tied to the process that Interprets it. and it does not exist without that context. But we still talk about adding links to a net, or facts to a data-base: assuming that the Interpreter is fixed. But the Interpreter also represents knowledge, about how to use the other knowledge. And if this "meta-knowledge" is also aquired. then the interpreter may not be an unchanging entity. Interpreter: Declarative/Procedural vs. Explicit/Implicit TEIRESIAS Flavell Split personalities, Post-hypnotic suggestions HAM like vs EPAM like models Why is it useful to think about knowledge knowledge "objectively To begin with, let us examine the idea that Cognition is the processing of Oinf ormat ion>, some internally stored symbolic representation of facts about the real world. This pervasive view of intelligence has assimilated the architecture for computer memory associated with -Ciref vonNeum4o> a separate, relatively low speed, passive storage distinct from the processors (arithmetic, logical. I/O. etc. ) that manipulate it. This is the way we have come to think about and build computers, but it was from the beginning only a compromise forced by the relative cost of active (switching) elements and their interconnections. To some extent, the brain has probably realized similar compromises, due to similar constraints -Cpref vonneumsB>, but not necessarily as far as the complete isolation of memory from processor as in current computer architectures. ssNor is it necessarily true that these constraints still apply to computer hardware design. Micro-electronic techology has inverted the cost ratio of components and interconnections, and the idea of distributed memory is a natural consequence that has been pointed out by several authors. (■Clref suther77>; -Clref rgsmi th77> ) . ■* - Hubel and - Weisel, frog's eye machines as formal symbol manipulators Al Insights: Logic, Production Systems - explict vs. objective knowledge -> symbolic "representation f <BARR>MK2. I representation: A* la CS 365. Tue 6-Feb-79 9: 4IPM knowledge as an object in the head .mark Objective Knowledge as Ascription What's really going on here? What do we mean when we say that someone knows something? When we see people driving a car or playing chess, we say that they "know how," and we interpret this in terms of their "having" the But the behaviors we've actually observed, like requisite knowledge. learning, remembering, making solving problems, and task, crudely explained by the are and are complex only performing saying that the knower "has" the requisite "knowledge" in his head. It is the way we talk and think about people that leads to the "knowledge as object" metaphor. Just as we say people have a love in their heart or have a pain in their neck, we objectivize the fact that they know how to do something and stick it in their head. -Ciref mccar77> refers to this as "ascription of knowledge to others. " . glitch Must . endg 1 itch check the McCarthy paper. In the next section I will discuss some alternative metaphors But I am not proposing that these metaphors for understanding knowledge. are in any sense more useful. In the remainder of the paper will propose the extension of an "objective knowledge" description to the The utility of "interpretive" parts of our models of human cognition. this metaphor stems from our limitations in understanding complex systems. It should be used as a tool for describing those systems, rather than being an assumption about the way those systems are. stoptext - language is "objective" why is it useful to think about X this way problems with "objective" knowledge" no evidence for facts in the head ( dec larati vely ) assume sufficient interpreter, how do we use the evidence no for the world out there, besides knowledge startte x t .mark Other Metaphors for Knowledge The Problem I am concerned with the characterization of Cognition and In particular, there is a generally accepted idea Cognitive processes. of what memory is in such systems, and I have poked at that notion a bit and found it unsatisfying. To wit, Memory in Man and Machine is a storage device of some sort, and the Cognitive processes have STORE and FETCH access to it. I <BARR>MK2. I Tue 6-Feb-79 9: 4IPM This is the basic yon Neumann idea of machine architecture. Our machines are still built that way. Current theories in Information Processing Psychology assume that that is the way the mind works. Psycholgical always assumption, theories haven't made that nor are computers necessarily going to always be built that way. SR theory, the rat in a maze Gestalt theory, gestalt closure Neuronal development Architecture, distributed memory, self-organization Compare this conception of memory as a storage device with that implied by a stimulus-response view of behavior. Here memory (of the appropriate response and the conditions that make appropriate) is viewed as a more-or-less structural change in the organism. Learning not only changes the ■C+contents of memory>, but the whole structure of the organism's it cognitive processes: An animal that remembers the experience of an electric shock in a T-maze context will seem to <+see> and 01earn> differently on returning to the maze than an untrained animal. Clearly this is an oversimplification of Behaviourism, but the important point is that it is always possible to view the acquisition of knowledge as a modification to a program. as opposed to an addition to a data base that is used by an unchanging interpreter. - memory biology, behavior, process structural change in the organism S-R psychology knowledge as procedures how can you know a process construction of reality Most everyone agrees that to some extent, we construct our personal reality. There are a lot of ways of looking at this issue: psychological, anthrop log ical, sociological, linguistic, philosophical, etc. The way that it relates to the material presented here is that if we give up the notion, that Memory is a passive storage device, and instead view it as an aspect of Cognitive behavior that is a constantly changing process, we have not evidence that there is a fixed external world that we record information about. / <BARR>MK2. I Conclusions about Tue 6-Feb-79 9: 4IPM knowledge ?The proceding discussion of the nature of knowledge is certainly naive, and may very well be wrong. Before going on the meta-knowledge. conclude, I would like to and I don't believe this is wrong, that the conception of knowledge as an accumulation of facts is only one way of Furthermore, that although it has served well looking at knowledge. as an epistomolgy for (classical) physics, for instance, to proceed in the study of cognition we will find other conceptions of the nature of knowledge useful. maybe why? just treat obj. k. differently? / <BARR>MK2. I Tue 6-Feb-79 9: 4IPM .mark Meta-knowledge And it is also possible to view an interpreting program in terms knowledge about its data base, called --:+meta~knowledge> in Al these days. Specifically, the interpreter must in some sense" know the extent and limit of available knowledge, what facts are relevant in a given situation, how dependable they are, and how they are to be used. This kind of knowledge is usually built implicitly into the interpreter of an Al system, but can be separated out and declared explicitly, leaving a residual interpreter whose job is to operate on the meta-knowledge. This duality between meta-knowledge and interpreting process has come to the forefront in Al Representation Theory through the development of representation languages which allow explict and flexible statements of meta-knowledge in their semantics. If all of the interpreter's knowledge, formerly encoded in implicit, procedural is made explict, as meta-knowledge, the interpretive of "code" that remains which, is relatively trivial, something like a scheduler of processes in turn, include meta-level interpretive information in their data. . glitch I have become familiar with the explicit use of meta-knowledge through Terry Winograd and KRL ({lref bobr77>). The best treatment of the relationship of representation and interpreting process is Brian Smith's recent work on Knowledge Representation Semantics (-Clref bsmith77>) I will return to this discussion in Section -(.'yon krl>. endg 1 i tch . * Thus, meta-knowledge and interpretive process can be viewed as playing the same role in Cognition: specifying how to find the relevant knowledge and how to use it. The distinction drawn between is like the familiar distinction between declarative and procedural knowledge representations, only one level up: The interpreter is procedural knowledge about how to reason with the base level knowledge, and meta-knowledge is its declarative form (see -Clref wino And when we consider our cognitive processes as declarative knowledge, meta-knowledge, it is natural to ask how we acquire it: do we learn to reason, and to remember? With this Duality of declarative meta-knowledge and interpreting process mind, it is fruitful to explore an epistemology that does not postulate a real, <+ob jec t i ve> world which we observe and store facts about. For if indeed our observations are recorded not as facts in a passive memory, but in the very process that is supposed to be doing the observing, the Oreal worldsmay be a simplification that can lead to confusions in the study of Cognition. An alternative epistemology, which I learned of through the uiork of the biol og i st/cy berneti ci st Humberto Maturana, has served as the inspiration for my own exploration of the nature of process in The name C+radical construe tivism> has been used by Cognition. ■Ciref glas77> to describe this view. Maturana recognizes the need } ' <BARR>MK2. KNOWLEDGE; I Tue 6-F'eb-79 9: 41PM to view memory as a structural change in the organism, but doesn't differentiate between the influencea of the external world and those induced by other states of nervous system activity. From this C+closure> property of the nervous system, he develops a theory of consciousness, including a formal description of the Observer. - relation to (identy with) cogntive process (MXI) explicit form of implicit knowledge in an interpretive process syntactic vs. semantic MX Consider the well-known human behavior called the t ip-of-the-tongue phenomenon: . crown ( 10, 10, 10) You run into someone you have met once and you can't remember his name. You remember very well your first meeting at a Mew Year's Eve party in and that he is the brother-in-law of your wife's boss. Then you remember that he has a foreign-sounding name. It rhymes spaghetti. with .endcrown skip continue You could use all of this knowledge in trying to recollect his name. And you would certainly say that you "know his name", even thought you can't recall it: You "know that you know it. " introspective This kind of evidence suggests a major role for Ometa-knowl ed ge> in human reasoning: Knowledge about the extent, reliability. importance, and history of the knowledge we "have" of the world. glitch Better fix up this to be clear that MX is a way of describing this behavior. Refs: (-Clref col 1 ins !mk >, -{lref teiresias>). col 1 ins ! gentner} and -Clref flavell}). teiresias.}, -Clref krl}, -Clref brachman}). endg l i tch (■Clref ("Clref . 1 - <BARR>MK2. I Tue 6-Feb-79 9: 4IPM Here is a partial list of the kinds of things that might be useful to know, about what we know: . crown (5, 5, 5) «Extent>. In particular, knowledge about what we don't know. To answer the question <+What is Paul Newman's telephone number?> you might reason that "If I knew Paul Newman's number, I would know that I knew it, because it is a notable fact. " «Defaults>. We have expectations about objects and events that we encounter that can be viewed as default values for the parameters of new data structures. How important some facts are relative to «Cr i ter ia 1 ity >. a given conclusion: A trunk is more indicative of an elephant than are four legs. Explicit knowledge about what kinds of knowledge «Relevance>. may be relevant to certain tasks. This may come in the form of links between data structures, or groupings of structures into larger structures. ■".XHistory>. Knowledge about the source, age, and development structure. of the data OCRel iabil ity>. How certain this knowledge is. If it is contradicted by other observations and when it should be abandoned. "C-CRepresentat ion>. Knowledge of the representation since it is necessary before the data structures can be modified. But explicit information about the representation is also possible, and it is especially useful for multiple uses of the same data structures. formalism is always available implicitly, <<Meta-cogni t ive knowledges. This is knowledge of the capabilities of the system. We humans learn this kind of information from experience {pref flavell}, but Al systems might also benefit from knowledge of their own performance. As an example, -Ciref bobr7s> discusses a robot who is planning a trip. His knowledge that he can read the street signs along the way to find is meta-cognitive knowledge. . endcrown out where he is <EARR>MK2. I Tue 6~Feb-79 9: 4IPM ,j How has it been done? what form has - examples from - KRL, MX taken in Al systems Al systems McCarthy, FOL Brian Smith The people Brian mentioned Collins: Al and Learning Systems, Reasoning from Incomplete X psych feeling of knowing Flavell's meta-cognitive knowledge Collins & Gentner what if any unifying design principles can be discovered Why is meta-knowlege useful - knowledge about knowledge and the interpretive process: an interpreter is "implicit" knowledge about the things it interprets. sometimes it is useful to have "explicit" knowledge about the "database". this is meta-knowledge (see Winograd's "procedural-declarative controversy") in fact it is essential for certain classes of activities, how can we characterize this class? Problems with MX - the problem of circularity, what is meta to what MX and Memory Critique and the search for a uniform representation