Lecture 3: Frame Problem

Program:

  • Frame problem "original", or inertia problme
  • Broader frame problem, holism problem
  • Phil. justificatoin-
  • Isotropy and being Quineian vs abduction proper and IBE
  • Fodor's pessimism and Thagard's optimism
  • Fridgeons, grue, bleen, cognitive wheels
  • Importance for XAI and general AI.

Nine dots puzzle

4 line solution:

3 line solution: requires background knowledge about that dots (no points!) have an area surface, so you can draw lines that meet, because they can go through the dots without having to be parallel. So the solution depends on our understanding of the representation of this problem.

9 points cannot be connected with three lines, 9 dots can.

Frame problem proper as inertia problem

We do not only need to represent what changes because of an event, but also what does not change. This is necessary for tractable and correct reasoning.

Inertia problem: Decide for all possible implications there relevancy -> intractability.

Humans can decide what is irrelevant without deliberate reasoning at all, it seems.

This is a big problem of logic: you'd have to specify all exceptions / irrelevant facts in a given situation.

This limited notion of the frame problem (in its computation/engineering sense) can be solved.

Solution: non-monotonic with negation as failure. Everything that is not explicitly mentioned as an exception, is not an exception.

Why is this non-monotic?

By adding knowledge "exception(pingu)" -> you can deduce less.

Frame problem within AI vs Cognitive Science

However, when we use these processes to describe human reasoning we encounter issues.

Q: How relevant is non-monotonic logic for studying cognition?

A: not much. Humans do not use these rules to derive the irrelevance of all kinds of salient perceptual inputs. We simply ignore them; they are not part of the deliberation.

We do not consciously derive all side-effects, also not by negation as failure (which is still a reasoning step).

So non-monotonic logic is a cognitive wheel: super handy from an engineering perspective; and yet, there are little examples of biological "wheels" in nature. A.k.a. the engineering solution does not help describing cognition. This was however an intent of GOFAI.

Dennett: all current solutions to the frame problem "proper" are cognitive wheels. Wheels are valid technical solutions, but not biologically plausible and useful for describing cognition.

Cf. McDermot (1987) argues the frame problem is not an AI problem precisely because the engineering part has been solved. The issue is dragged into AI by philosophers.

McDermot's Sleeping dog strategy.

Holistic version of the frame problem

Fodor, also in 1987: AI is not bothered by frame problem only because they are stupid.

Q: is AI still stupid nowadays?

E.g. data-driven chatbots that become racist?

Fodor's Fridgeons

Versus sleeping dog strategy.

Sleeping dog strategy: unless specified, things remain the same. Like in the negation by failure.

This makes one big assumption: namely that most things do not change given an action.

Fodor's point: we cannot distinguish when or when not this assumption holds.

Counter: putting on Fodor's fridge renames every particle as a fridgeon.

The crux: you need an ontology to show that the previous "fridgeons" do not make sense. How can we metaphysically exclude fridgeons? How can we choose the right ontology?

Cf. Dennett's heterophenomenology. What must we know?

Problem of holism

Haselager: "Seeing the relevant consequences of an event made possible by an understanding of the situation by using what one knows.

  • How are relevant pieces of knowledge found?
  • How do they influence one's understanding of the situation?
    • i.e. how they guide the generation of explanatory hypotheses and bear their acceptibility

So the frame problem is relevant for non-demonstrative reasoning.

The sleeping dog strategy solves the limited form of the frame problem, but moves to another problem:

Also question is not to "solve" the frame problem. We have solved it: humans know what is relevant and what is not. The point is to understand it so we can apply it to AI.

Two approaches

1) Representational approach 2) Inferential approach: lack of understanding; change it into a problem about ontologies. * Criticism: most "solutions" are just off-loading the problem, e.g. to "evolution". * We cannot learn everything, aha, but evolution took care of that. * Fodor's criticism: This does not give fundamental insight in non-demonstrative inference.

Isotropy and Being Quineian

Argues that these two are elements of human cognition:

Isotropy: everything may be relevant for everything (can be) Related to abduction proper. E.g. butterfly effect.

Q. I had a question about the Raven paradox: is isometry somewhat similar in the sense that any observation could support a hypothesis? (I would not argue no).

Quineian: degree of confirmation beliefs on overall belief system. Related to IBE.

Significance of Frame problem w.r.t. non-demonstrative inference

FP computational version of difficulties surrounding non-demonstrative inference as found in induction, abduction and IBE.

Isotropy and being Quineian are part of successful human inferences, but we cannot explain how it works and that we can implement in a AI. AI as proof of concept: if we would truly understand it, we would be able to implement it.

Goodman's new riddle on induction

X is grue iff X is green before 1 Jan 2020 and blue afterwards.

Y is bleen iff Y is blue before 1 Jan 2020 and green afterwards.

We are able to reason this way; e.g. change from unborn infant to baby. But we don't want these constructions in everyday reasoning.

If we observe a green emerald we have no less inductive validity to call it grue rather than green (?) REVIEW.

There is no metaphysical inductive justification.

Why is grue a good predicate and green not? The issue is not complexity.

Psychological plausibility

Can we tractably do isotropic and Quineian inference?

Cf. P vs NP -> we lack key mathematical notions.

Thagard's optimism: + sub-symbolic approaches, spread of activation. Coherence is NP-hard.

Relevance of Frame problem of XAI

What is relevance for:

  • explainable AI
    • What is the human likely to find relevant?
  • general AI?
  • practical AI?
    • Maybe not here
  • cognitive science?
    • Kwisthout: important, how can we make sense with this overload of information tractably? (broader FP)
  • neuroscience?
    • For unifying theories of cognition such as predictive coding (brain minimizes prediction errors)
    • Dealing with so much input (and equally much prediction errors), we need a relevance notion