We use cookies to distinguish you from other users and to provide you with a better experience on our websites. Close this message to accept cookies or find out how to manage your cookie settings.
To save content items to your account,
please confirm that you agree to abide by our usage policies.
If this is the first time you use this feature, you will be asked to authorise Cambridge Core to connect with your account.
Find out more about saving content to .
To save content items to your Kindle, first ensure no-reply@cambridge.org
is added to your Approved Personal Document E-mail List under your Personal Document Settings
on the Manage Your Content and Devices page of your Amazon account. Then enter the ‘name’ part
of your Kindle email address below.
Find out more about saving to your Kindle.
Note you can select to save to either the @free.kindle.com or @kindle.com variations.
‘@free.kindle.com’ emails are free but can only be saved to your device when it is connected to wi-fi.
‘@kindle.com’ emails can be delivered even when you are not connected to wi-fi, but note that service fees apply.
I am writing this paper only for those people who will agree with me that research in AI very often lacks a disciplined approach and that this situation should be changed. Trying to establish this fact for those who disagree with me and to do so in a way acceptable to a normal scientific community will need much more work than I can afford and, if what I say is true, would be a futile thing to do anyway.
Of course, even though one may agree with me in my above view, one may not agree with me about how the situation should be changed. My fear is that the situation will be hard to change: it will not happen till a number of years have elapsed after there is a change in our approach to the field of computer science.
Even to argue that point I will have to assume that AI depends heavily on computers and on programming. I think that it is safe to believe that we have agreement on that. If there are any proponents of “disembodied AI,” I might be one among them: and I am quite prepared to keep that view in abeyance.
This section is the least concerned with concrete methodological issues, but there is, we believe, a need to explore some of the arguments concerning the abstract representations that underlie programs and theories in AI. Several other collections of philosophically-oriented AI papers address these issues quite well, and these are covered in the annotated bibliography. Nevertheless, there are a number of arguments that are either new, or will contribute significantly to the unity of this particular collection.
Winograd addresses, at a very general level, the promises and problems of AI: he is no longer as optimistic as he used to be. He briefly reviews the physical symbol system hypothesis (PSSH, see Newell, 1980) which is typically taken as one of the foundation stones of AI. Winograd then singles out for attack the essential sub-component, representation – “complete and systematic representation is crucial to the paradigm.”
He claims that an underlying philosophy of “patchwork rationalism” is responsible for the unwarranted optimism that pervades much of AI. And he briefly introduces an alternative basis – “hermeneutic constructivism.”
He examines three types of problem that arise from the misguided foundations of AI: gaps of anticipation, blindness of representation, and restriction of the domain. He sees the PSSH as leading AI into the necessity for decontextualized meaning – i.e. AI systems are constructed with rules that “deal only with the symbols, not their interpretation.” We are forced to deal with the representation as an object “stripped of open-ended ambiguities and shadings.” Only a very limited form of AI can be constructed on this basis, Winograd claims.
This discussion seeks to compare, in very brief compass, a current radical argument for connectionism and a radical argument against. It is not clear that the very same version of connectionism is defended by Smolensky as is attacked by Fodor, but since I do not bring the two arguments directly in contact, that will not matter. My own inconclusive view is that the jury is still out, and there is no pressing need at the moment to believe what Smolensky says, though one may respect it and be stimulated by it, nor to reject the whole thing on the grounds Fodor gives. One can legitimately be, in a narrow and strict sense, an agnostic, without giving that word the force of active disbelief it is often made to carry.
Smolensky
Smolensky (this volume) declines to base his version of connectionism on the brain: on the supposed analogy of the brain to connectionist networks, and for this all critics must give much thanks. Connectionism for him must stand or fall on its own merits as a model of processing and behavior, and not on gross and randomly drawn similarities with natural phenomena, similarities that cannot be assessed in any scientific manner, but which either appeal or fail to do so, depending on the mood of the audience.
The rationalist philosophical tradition which descends from Socrates, to Plato, to Leibniz, to Kant, to conventional AI and knowledge engineering takes it for granted that understanding a domain consists in having a theory about that domain. A theory formulates the relationships between objective, context-free features (attributes, factors, data points, cues, etc.) in terms of abstract principles (covering laws, rules, programs, etc.) As this tradition develops, even everyday practice is assumed to be based on unconscious theory.
In one of his earliest dialogues, The Euthyphro, Plato tells us of an encounter between Socrates and Euthyphro, a religious prophet and so an expert on pious behavior. Socrates asks Euthyphro to tell him how to recognize piety: “I want to know what is characteristic of piety … to use as a standard whereby to judge your actions and those of other men.” But instead of revealing his piety-recognizing principles, Euthyphro does just what every expert does when cornered by Socrates. He gives him examples from his field of expertise; in this case situations in the past in which men and gods have done things which everyone considers pious. Socrates persists throughout the dialogue in demanding that Euthyphro tell him his rules, but although Euthyphro claims he knows how to tell pious acts from impious ones, he cannot state the rules which generate his judgments. Socrates ran into the same problem with craftsmen, poets, and even statesmen. None could articulate the theory underlying his behavior.
I want to ask ‘What kind of field is artificial intelligence?’ and to give an answer. Why is this an important question? Because there is evidence within AI of a methodological malaise, and part of the reason for this malaise is that there is not a generally agreed answer to the question.
As an illustration, several papers in this volume draw a number of different analogies between artificial intelligence and other fields. AI is compared to physics, to chemical engineering, to thermodynamics and to psychology; in fact it is said to be psychology. Each of these is a very different kind of field with different kinds of methodology, criteria for assessing research and so on. Depending on which of these you think artificial intelligence is really like, you would decide what to do, how to do it, and how to assess other people's work.
Evidence of malaise
One of the symptoms of this malaise is a difference amongst referees of papers as to the standard which is expected for conferences, journals, etc. When I was programme chairman of a major AI conference, I noted that for more than 50 percent of the papers the referees disagreed as to whether the papers should be accepted or rejected. And this wasn't just a question of having different thresholds of acceptability, because the opinions would reverse on other kinds of papers. So clearly the referees were applying very different criteria when deciding which papers were worth accepting.
A continuing AI debate is the relationship of programs to theories. Some clarification of this issue would do much to put AI on firmer foundations.
Wilks attempts to clarify some basic issues concerning the relationships between theories and models by laying out a taxonomy for these two variously-used terms. Using the example of AI programs for language processing, he argues that such programs are not so much theories in the more classical senses, but can usefully be considered to be theories of computational processes.
We have reprinted the Bundy and Ohlsson debate from the AISB Quarterly (1984). In this interchange of ideas as to what one has a right to expect from AI research, and, in particular, what the relationship between programs and theories ought ideally to be, we see apparently very different opinions move surprisingly close together once certain terms (such as ‘principle’) have been clarified. Nevertheless, Bundy remains committed to the significance of mechanisms or techniques as the essential principles in an AI program; and Ohlsson stands by his demand for validating behavioral principles as the important product of AI programming. These two viewpoints correspond to the representations at the two ends of the sequence of abstractions discussed by Partridge in the previous chapter. And, interestingly, Bundy's analysis of Ohlsson's principle bogs down at several points just because the meaning of seemingly straightforward terms in the principle are in fact open to several interpretations and there are no more-precise restatements to fall back on (as there would be if the principle existed in the context of a sequence of abstractions).
What I want to sell you is some neuroscience. I want to sell you the idea that current neuroscience has become directly relevant to the cognitive/computational issues that have always concerned AI. In what follows I shall outline a highly general scheme for representation and computation, a scheme inspired in part by the microarchitecture of both the cerebral and the cerebellar cortex. We shall also explore some of the interesting properties displayed by computing systems of this general kind.
The basic claim, to be explained as we proceed, is that the brain represents specific states of affairs by implementing specific positions in an appropriate state space, and it performs computations on such representations by means of general coordinate transformations from one state space to another. (There is no suggestion that this is the only mode of information-processing in the brain. But its role does seem to be nontrivial.)
That one can perform computations by this abstract means will be old news to some readers. For them, I hope to occasion surprise with evidence about the implementation of these procedures in the microanatomy of the empirical brain. Also intriguing is the natural way in which problems of sensorimotor coordination can be solved by this approach, since from an evolutionary point of view, sensorimotor coordination is where cognitive activity had its raw beginnings. Of further interest is the very great representational power of such systems, and the truly extraordinary speed with which even biological implementations of such procedures can perform computations of great complexity.
If there are defining characteristics of AI, one must be the centrality of programs. Computer science or software engineering might be considered to have a better claim to this characteristic, but closer inspection of the role of programs within those fields reveals that not to be so. Formal specifications and means of realizing them in machine-executable representations dominate these fields; individual programs are relatively secondary in importance. But in AI implementations (programs) have always played the central role.
Bundy offers a three-fold division of AI, and claims that the role of a program is dependent upon which subdivision some particular AI work falls within. His three categories of AI are “applied AI,” “cognitive science,” and “basic AI.” It is important to sort out these issues (he claims) if we are to cure the “methodological malaise” in AI.
Dietrich then argues that a keystone in the foundations of modern AI is that intelligent thought is symbol manipulation, and, moreover, symbol manipulation of the type manifest in Turing-von-Neumann computers. Hence, the search for AI is reduced to finding the right program, but with no guide as to how to find it.
The role of programs in AI (according to Dietrich) is little better than random attempts to hit upon this program (or class of algorithms) whose existence (but little else) is promised by the foundational theory.
Most of the difficulty of including AI in the standard collection of sciences is that there are recurring features of the best or most-publicized work in AI that are hard to fit into the conventional observation-hypothesis-deduction-observation pattern of those sciences. Some parts of the difficulty can be clarified and resolved by showing that certain details of what is involved in AI have underlying similarities with corresponding steps in other sciences, despite their superficial differences. The treatment of ‘theories’ and ‘programs’ below is intended as a commentary on that remark. A further part of the difficulty is that AI still lacks some scientific credibility because it does not yet seem to have the standard of reproducibility and communicability of results that is built into other sciences. The difficulty is illuminated by activities in AI research that have come to be known as ‘rational reconstructions’. While the earliest attempts at rational reconstruction have generally been less than successful, the idea itself is potentially useful as a means of generating new knowledge or mapping out new territory in AI.
On theories, models and representations
Sciences are supposed to have underlying theories, and technologies rely on theories through the help of their supporting sciences. The first hesitation among outside observers to give AI full credit for being a science or technology comes from the difficulty of identifying AI's theory or theories.
Of course, unless one has a theory, one cannot expect much help from a computer (unless it has a theory).
Marvin Minsky
Introduction
Computer programs play no single role in artificial intelligence. To some, programs are an end; to others, they are a means. These two groups might be thought to contain those who think AI is an engineering discipline, and those who think AI is a science. This is only partially true; the real situation is more complicated.
The first group is by far the largest and contains many of the most prominent AI researchers. For example, in his book Problem Solving Methods in Artificial Intelligence, Nils Nilsson states that
Future progress in [artificial intelligence] will depend on the development of both practical and theoretical knowledge … As regards theoretical knowledge, some have sought a unified theory of artificial intelligence. My view is that artificial intelligence is (or soon will be) an engineering discipline since its primary goal is to build things. (1971, pp. vii-viii, his emphasis)
Barr and Feigenbaum (taking a slightly more cautious position) also claim that “whether or not [AI leads] to a better understanding of the mind, there is every evidence that [AI] will lead to a new intelligent technology” (1981/1982, p. 3, their emphasis).
Many researchers who see themselves as theorists or scientists also belong in this group because they think that the ultimate goal of their work on theory is to produce a computer program that does something useful, whereas in other disciplines, the goal of theorists and scientists is to produce a theory.
Commercial AI is now big business, but AI itself is full of fundamental, unsolved problems; so what exactly is being marketed? Whether we wish to call it AI or expert systems, what are the scope and limitations of this technology?
Hewitt argues that there are severe limitations inherent in the ‘logic programming’ movement, because of its close association with current expert systems technology. He sees a need to deal with “open” systems that involve inconsistent knowledge and will need “due process reasoning” for decision-making.
On a broader front, Hewitt is questioning the suitability of the symbolicsearch- space paradigm (the foundational hypothesis of AI for the last thirty years) as a basis for intelligent systems. This paradigm, with its prerequisites of well-defined initial states, goal states, and state-transformation operators, is applicable to “artificial domains like chess and mathematical theorem proving. It is not very adaptable to the hurly-burly of solving problems involving interaction with the physical world.”
The Dreyfus brothers also argue that current expert systems' technology (CEST) is severely limited and it is built upon a fundamentally misguided view of human expertise. They see future progress in AI only when some fundamental assumptions are abandoned (such as knowledge as a collection of context-free units – another manifestation of the general pervasiveness of the symbolic-search-space paradigm) and radically new approaches to the problem are taken, such as attempting to understand and implement, on appropriate architecture, holistic reasoning methods.
The state of the art in any science includes the criteria for evaluating research. Like every other aspect of science, it has to be developed. The criteria for evaluating AI research are not in very good shape. If we had better standards for evaluating research results in AI the field would progress faster.
One problem we have yet to overcome might be called the “Look, ma, no hands” syndrome. A paper reports that a computer has been programmed to do what no computer program has previously done, and that constitutes the report. How science has been advanced by this work or other people are aided in their research may not be apparent.
Some people put the problem in moral terms and accuse others of trying to fool the funding agencies and the public. However, there is no reason to suppose that people in AI are less motivated than other scientists to do good work. Indeed I have no information that the average quality of work in AI is less than that in other fields. I have grumbled about there being insufficient basic research, but one of the reasons for this is the difficulty of evaluating whether a piece of research has made basic progress.
It seems that evaluation should be based on the kind of advance the research purports to be. I haven't been able to develop a complete set of criteria, but here are some considerations.
Paradigmatic confusion in AI. In spite of what I regard as AI's significant achievements in beginning to provide a computational language to talk about the nature of intelligence, the not so well-kept secret is that AI is internally in a paradigmatic mess. There is really no broad agreement on the essential nature or formal basis of intelligence and the proper theoretical framework for it.
Intelligence as information processing on representations
Let us first seek some unities. There is something that is shared almost universally among workers in AI: “Significant (all?) aspects of cognition and perception are best understood/modeled as information processing activities on representations.” The dominant tradition within AI has been the symbolic paradigm. On the other hand, modern connectionists (and the earlier perceptron theorists) offer largely analog processes implemented by weights of connections in a network. Stronger versions of the symbolic paradigm have been proposed by Newell as the physical symbol system hypothesis (Newell, 1980), and elaborated by Pylyshyn (1984) in his thesis that computation is not simply a metaphorical language to talk about cognition, but that cognition is literally computation over symbol systems. It is important to emphasize that this thesis does not imply a belief in the practical sufficiency of current von Neuman computers for the task, or a restriction to serial computation. Often, disagreements with the symbolic paradigm turn out to be arguments for parallel computers of some type, rather than arguments against computations on discrete symbolic representations.
There are few principles on which nearly all practitioners of AI will agree. One of them is that intelligence is the formal manipulations of symbols. This nearly unanimous consensus is being systematically challenged by an approach to AI that models intelligence as the passing of numerical activation values within a large network of simple parallel processors. In this connectionist approach to AI, intelligence is an emergent property of the network's processing: each individual processor has no intelligence, and the messages they exchange – real numbers – participate only in very simple numerical operations. Input to the network is coded as a set of numerical activity values on the input processors, and after this activity has propagated through the connections in the network, a pattern of activity appears on the output processors: this pattern encodes the system's output for the given input. Each connection between processing units has a numerical strength or weight, each unit typically computes its activity by using these weights to form the weighted sum of the activity of all the units giving it input, and passing this weighted sum through a non-linear response function such as a threshold or sigmoid curve.
This paper addresses the sense in which intelligence is supposed to “emerge” in these connectionist systems, and the relation that this implies between the connectionist and traditional approaches to AI. The characterization I will formulate for a connectionist approach to AI is controversial in certain respects, and is not intended as a consensus connectionist view.
Needham sets out the view that AI has nothing special about it (except perhaps an unhealthy concern with being special). He compares AI to chemical engineering and finds that there are no real differences, and concludes that we should get on with what is essentially engineering and cease worrying about ‘foundations.’
Sparck Jones explores the nature of programs as experiments in AI, and comes to the conclusion that if we view AI programming as an engineering type of enterprise, we obtain an appropriate interpretation of events. In particular, she examines the idea of adequacy of AI programs, which is rather different from the notions of correctness that software engineers normally associate with their programming endeavours.
Finally, we have reprinted McCarthy's Presidential message to the American Association for AI (AAAI). He laments the lack of agreed standards for evaluating AI research. AI is special in the sense that it claims to be a science and yet the research methodology is undefined (or illdefined) in a number of critical places. In particular, he argues that we need standards for evaluating research results in AI.