Dreyfus Class Presentation 1/26/98

I. The philosophical foundation of Dreyfus' criticism of AI is found in the work of Heidegger, Merleau-Ponty, and (the later) Wittgenstein.

II. The (classical AI) tradition he is fighting goes way back in philosophy.
A. Socrates thinks that to understand (say) piety you need to know the rules for distinguishing the pious from the non-pious. But in the dialogues the disputants are unable to do so. As we would put it to day, we are unable to give the necessary and sufficient conditions for piety. Plato reaction to this is to believe that though we do not know the rules they are part of the unconscious equipment provided by the soul. Ideas we forgot when the soul became embodied. The role of philosophy is to recover these lost rules.
B. AI researchers are attempting to do the same. To find the rules by learning them or guessing at them in order to duplicate intelligence in rule-governed machines.
C. The fundamental ideas go way back

1. Hobbes: Thinking = Calculation = operating on the physical shapes of symbols (their syntax) to calculate new results.
   
2. Descartes: Understanding = analysis into simple elements (primitives). And the mind contains representations of the world formed from these elements.
   
3. Leibniz: All concepts can be represented by numbers. So all reasoning with concepts amounts to mathematical calculation. Even skills can be captured in this way.
   
4. Kant: Concepts are really rules operating on combinations of elements.
   
5. Husserl: The story about rules is much more complex. Now we need hierarchies of rules governing rules governing rules ... For example, to have the concept of (say) an eraser is to understand the set of all its possible features under any possible viewpoint or context. Husserl called this the eraser frame. The frame is a data structure listing all we need to know about erasers: their essential properties, their default (or typical) properties, etc..
   D. Marvin Minsky used exactly the idea of frames in AI and Doug Lenat's CYC project continues in these footsteps.
   
   

III. The philosophers in this classical tradition are generally rationalists. However, there have been other philosophers who have objected to these ideas:
A. Aristotle for example did not believe that intelligence or ethical knowledge could be formulated as rules.
B. Hume also objects to the idea that concepts can be expressed as rules.
C. Wittgenstein is a transitional figure. His early work (in the Tractatus ) accepts the classical view, but a lot of his later work (for example the Investigations ) criticizes his previous thought. In particular, he points out in the Investigations that there are no primitives in the world. What counts as the simple element depends on the context: your goals plans or purposes. You might break a chair down into its parts in very different ways depending on whether you were buying it, painting it, or shipping it. It is impossible to give rules for what counts as a chair. It's not something with legs and a back (beanbag chair) nor is it something you sit on (because then bicycle seats would count as chairs and they don't).
D. People in AI might reply. OK all we need to do is make rules for how the context affects our concepts explicit. But there are arguments that would suggest that this is a losing game.

1. Wittgenstein: The regress of rules problem. Suppose I have a rule for recognizing plants. Then I need the rules for recognizing the elements mentioned in this rule: leaf bark green oval but these rules will in turn need rules for recognizing their elements. This process cannot go on forever. If I apply the plant rule at all eventually I must come to items that i simply recognize without the need for any rules. But if the basic elements can be recognized without using any rules why not say that plants can be recognized without using any rules?
   
2. Heidegger: The problem of relevance. To deal intelligently with any problem we need to know what is relevant and what is not. Knowledge is not just knowing facts but knowing what facts are relevant in a given situation. Example: what is relevant to the performance of a jockey is his/her previous record except (say) if he/she is allergic and the track has goldenrod. Ordinarily goldenrod at the track has nothing to do with who will win, but our understanding of relevance lets us know the exception. It does if the jockey is allergic. Now can we give the rules for relevance? No because we need rules about relevance of the various relevance rules, and rules for these rules etc. Lenat's attempt to exit from the problem by explicitly writing relevance rules failed after over 1,000 rules were invented.