Cognitive wheels: the frame problem of AI | Daniel Dennett

Transcript of analog notes. If citing from here, make sure to look at the original text again.

What is the Frame Problem (FP from now on).

Coined in 1969 by McCarthy & Hayes, related to a planning problem ("frame problem proper"). It is nowadays understood as a broader problem that is not specific to a technical problem in AI, but applies to efforts to understand cognition and intelligence in general.

p. 131: Dualists evade but not solve the problem; the avoid answering the tough questions on how we think.

Intelligence can be understood as such: we think before we act, and our actions are guided by what we know (e.g. from experience).

The FP is also not visible for Hume and mentalistic philosophies because they operate on a phenomenological level. Dennett summarizes the phenomenological approach as follows: namely that items in view are individuated by their meaning.

On page 133 Dennett shows with examples how Hume's approach explains the FP away. By departing from "meaning" and "habits of expectation" we does not explain the intricacies of these habits. (And I guess also not where meaning comes from; it's just there. But I don't recall if Dennett says something like this).

Limits of introspection and the suggested improvement

Various implications or assumptions of actions are easily overlooked and do not seem learned from experience. This points to the limitations of the phenomenological approach to the FP.

Instead, we should do hetero-phenomenology, which fits with the typical AI-approach. Now we ask: what are the informational needs of a specific task? What must be known or done for a given magician's trick? (132)

AI forces also "banal information to the surface" (135).

An AI that starts as a tabula rasa is not sufficient. In AI however, you can "install" all knowledge needed by some set of tasks. The question then is: how do we install the appropriate knowledge into an AI?

Installing information

This includes a semantic and a syntactic issue (136) of how we can represent our knowledge so that it is usable.

The semantic aspect can be seen as the "backstage" of this problem, but this level is not accessible to introspection.

Given is: I don't observe myself planning.

Then there is two different flavours of phenomenology.

Vanilla phenomenology must then conclude that if I don't observe myself planning, there is no planning (it's not a phenomenon after all).

Instead, Dennett takes the hetero-phenomenological approach by saying: there must be planning, otherwise our behavior would be magical! (137) Otherwise we simply couldn't meet the informational demands of the performed task. Note that this ascribes something mental without introspection, but instead based on the informational needs of a task.

So far our understanding of planning saw it as a conscious process. But how then, does the presumed unconscious use and testing of background information in everyday planning work?

The desired behavior for an AI here should meet the intelligence test of counterfactuals. We would require an AI / robot to act "surprised" when its expectations are thwarted. (139)

The syntactic problem has several aspects (136):

  • There is a need for a generative or productive system of representing information
  • A Spinozistic deductive system is not sufficient for retrieving our knowledge (i.e. one with a manageable amount of axioms)
    • Instead we rely on experience to know how the world is
    • What presents itself in experience is not deductively reliant on what we knew about the world before.
  • It has to be memory efficient
  • It has to work timely; information has to be accessible in time, otherwise it's useless.
    • Being 'smart' is not just finding the solution; it is finding the solution when it still matters.

The FP can be described as the problem how to represent knowledge such that the above requirements are met.

The FP is not an induction problem in disguise

N.B. that in Haselager's summary of Fodor's position on the FP, Fodor actually does hold that ultimately the FP is a problem of non-demonstrative inference and requires key improvements in confirmational inductive logic.

The following is comprehensive and contains a definition of the FP:

The frame problem is not the problem of induction in disguise. For suppose the problem of induction were solved. Suppose - perhaps miraculously - that our agent has solved all its induction problems or had them solved by fiat; it believes, then, all the right generalizations from its evidence, and associates with all of them the appropriate probabilities and conditional probabilities. This agent, ex hypothesi, believes just what it ought to believe about all empirical matters in its ken, including the probabilities of future events. It might still have a bad case of the framing problem, for that problem concerns how to represent (so it can be used) all that hard-won empirical information - a problem that arises independently of the truth value, probability, warranted assertability, or subjective certainty of any of it. Even if you have excellent knowledge (and not mere belief) about the changing world, how can this knowledge be represented so that it can be efficaciously brought to bear? (140)

Note that in Haselager's summary, Dennett's stance on the FP would be in the non-deductivist camp, and then in the "representational" subgroup, i.e. solving the frame problem is mainly about finding the right kind of structure to represent knowledge used in common-sense reasoning.

The deductive approach to the FP

Frame axioms, p. 141 are background axioms "which describe general conditions and the general effects of every action type defined for the system." (141)

This includes no-change axioms.

Minsky: axioms in a (monotonic) logical system always "allow" some new inference to be drawn; that is they necessarily add something to the knowledge base. Another axiom implies more theorems. In monotonic logic there is thus no way to make a rule about what conclusions should not be drawn. Axioms about relevancy do not avoid producing the unwanted theorems as well; you only add extra statements about irrelevancy.

In human reasoning, endless lists of exceptions are avoided with a simple ceteris paribus clause. But sometimes this assumption turns out to be incorrect. By adjusting our assumptions with extra theorems, some of the things that were true in the previous logical system no longer hold. In other words, adding axioms can make things that were true before untrue. I.e. to capture human reasoning we minimally need a non-monotonic logic.

In the deductivist approach to FP, the FP appears as a qualification problem (by McCarthy): How to make a system ignore what it ought to ignore? (143) N.B. elsewhere this is named the inertia problem.

Proposed solutions to the qualification problem

p. 144-145:

  • Stereotypes, frames, scripts
  • non-monotonic logic (McDermott, Doyule)
  • default reasoning (Reiter)
  • circumscription (McCarthy)

Cognitive wheels

p. 146.

A cognitive wheel is a design proposal in cognitive theory that is clever but is profoundly biologically unplausible.

The big question here is: Is AI just a toolbox of cognitive wheels?

Dennett says that despite the unbiological implementation there may still be fruitful isomorphisms with biology on a higher level of description (cf. Marr's levels).

The takeaway here is that when we discuss whether AI is a cognitive wheel, we should not be sceptical a priori but instead ask: on what level of description are we talking?

On the implementational level AI may be wildly biologically unplausible; but this may not be so on the computational level.

Of course, the question then is: how do we link phenomenological and implementational levels of description? (148)