TFCA Guest Lecture 11 by Mark Blokpoel on Analogical Inference

Discussion of Analogical Inference .

Inspiration: Blokpoel's PhD was on abductive inference, inspired by studies on social communication... e.g. on action understanding. What's new is to take a formalized approach.

Can Bayesian models be used to explain communication? This attempt got stuck a bit, because it's not always clear what language can be used to communicate (this is assumed in the Bayesian framework, I guess). In Bayesian models the hypothesis space is assumed.

So there was a prior question: how to you establish a new form of communication? How do you figure out which form of communication is best?

Full nativism vs production

Where do hypothesis come from?

  1. Full nativist
    • All hypothesis that we can have in principle are naturally hardwired
    • Jerry Fodor's position
    • Fodor's motivation came from the computational limitations of explanatory frameworks
    • Apparently some people using Bayes for cognitive modeling are also mildly nativist.
  2. Production/generative
    • Do not assume the hypothesis space is built in and can be generated on the fly during the lifetime of an individual

Abductive Inference

Complete Abductive Inference:

Input: evidence e, knowledge K Output: Inference to the best Explanation(e, abduction proper(e,K))

If you can assume that H is built in, then Abduction Proper collapses to IBE: i.e. the whole story the is "Inference to the best Explanation(e, H)".

Example of SMT

Theory about how people make analogical inferences.

Bohr: electron revolution around the nucleus can be thought of as an analogy of how planets revolve around the sun.

The idea here is that this analogism helped Bohr generate his explanation of how electrons revolve around a nucleus.

Have a relational predicate structure of the solar system model. Objects at the bottom, higher up attributes, higher up relations between attributes or objects.

E.g. "greater" (see slides) are of same type (binary relation) and have same label ("greater").

Parallel connectivity: anything that's directly? below needs to match as well. Which is true, because the attribute types match, but their label differ.

Potential issue with this match: mapping "causes" requires "gravity" and "opposite" to match, but they are relations with a different label.

Projection: if you have matching structures, then connecting labels not in the one domain, may be transferred into the other.

Systematicity: Once you find a match, the quality is measured by the depth of its structure. E.g. systematicity 3 in the example in the slides.

Maybe systematicity relates to the prior probability of a hypothesis. (See question below).

Augmentation: Given an original representation, you can add different pieces onto it using projection with matches in the knowledge base.

You do this process for each core in your knowledge base. (This is what satisfies isotropy; anything in the knowledge base may be used to generate hypotheses).

Knowledge core is a "concept" from knowledge base K.

Space of different interpretations? Paper calls it reinterpretations. But doesn't he mean representations, strictly speaking? Does interpretation not imply a meaning which is distinct from a representation.

Some remarks on the properties of abduction proper

They notes on the paper. Here I added some extra informal remarks if they added something.

I had a question on novelty and groundedness.

  • Isotropy:

In the field of communication this property is quite obvious, i.e. we can talk about anything we know at any given time; any knowledge we have may be relevant for understanding some communicative signal.

  • Groundedness:

Your abductive inference should be able to explain things you perceive.

What about internal states? I had a question about this, Mark said that they put this remark in.

In some cases abduction proper may explain internal states.

Discussion

  1. Question on being Quinean. Why is this not mentioned?

Blokpoel: correctness is a property of IBE, not abduction proper. Theory only concerns the generation.

My remark: does sensibility reintroduce the issue of correctness and perhaps that of Being Quinean?

Sensibility: sort of lower bound on correctness of the hypothesis. But in the capacity it is used of the paper (sort of holds for groundedness as well), it's not that Quinean.

  1. Does sensibility imply a sufficiently high prior probability?

Does 'being sensible' imply 'having a reasonable prior probability of being true'? For example (from Putnam, 1981) you might explain an apparent picture of Winston Churchill in the sand as the trace of an ant crawling on the beach. If that ant actually crawled in that particular way, leaving that particular trace, it would perfectly explain that you see this picture. This hypothesis is typically being dismissed on the ground that its prior probability is neglectibly low; hence, the posterior probability P(ant crawl = true | image) is very low, while the likelihood P(image | ant crawl = true) may be very high. Can you comment on that?

In the analogy model this very unlikely hypothesis would be part of the hypothesis space.

In a Bayesian approach, would the prior itself be part of the hypothesis space?

Follow-up: would it also be in the hypothesis set that you use in IBE?

The model currently does make claims about the concrete relation with IBE. The hypothesis generation procedure is now as general as it can be, so it is likely that it needs to be constrained further. Perhaps in that process we would also have to exclude very unlikely hypotheses.

The deepness of a match (e.g. systematicity) could perhaps also be a measure for determining likelihood? This is uncharted territory.

What are other attempts?

Churches and hierarchical Bayes (HB): they do provide an interface between hypothesis generation and actual (Bayesian) inference, that this model does not offer.

Why didn't you pursue that approach? Recap of HB: set of generative grammars, can generate variables and their values using Bayesian inference. Impressive example: they used HB to explain how we can explain and interpret handwritten symbols, by giving a set of generative rules to generate scribbles, and then inferring the most likely interpretation of those scribbles.

HB could in principle fulfill a similar role, but there do not seem to be any cognitive constraints on the generative grammars to be a model for abductive inference. It's more of a general framework.

They went with SMT because it is more specific about generation and has more empirical support.

My question: how novel is novel?

Does analogy allow for truly new insights? New candidate hypotheses are always related, by analogy, to an existing knowledge base K. Would this framework be able to accommodate the type of reconceptualizations that amount to a paradigm shift?

After formulating this question, I did see that the paper somewhat addresses my question, by pointing out that there is a shift from "the burden of explaining where candidate hypotheses come from to explaining where knowledge comes from." (9)

The answer to this anticipated comment is that "Knowledge acquisition is, however, a different explanatory target." (9)

I'm not entirely sure I get that. Isn't an important reason why we study abduction proper and IBE to understand how new scientific knowledge comes about?

This may be an interesting point to discuss, especially insofar as such a paradigm shift itself requires the formulation of a new hypothesis/core theory.

Another thing that I noticed in this context is that novelty is defined as a hypothesis that a person has not generated before. So novelty is defined with respect to a person rather than a knowledge base in more general terms. This may be a conceptual distinction that is relevant in discussing my question.

Answer 1: If you bring in new knowledge, you may radically change the hypothesis space, especially if you combine it with IBE. New knowledge -> new hypothesis.

Then the question is, where does the new knowledge come from? IBE.

Answer 2: There are already competing hypothesis in the hypothesis space, but due to background knowledge we do not actually infer one. A paradigm shift could be thought of as a shift from inferring from one hypothesis set to the other. So one answer is: a paradigm shift is within the hypothesis space (in that sense, not radically new).

Mark really liked my question and actually mentioned that Steve Levinson asked a nearly identical question at his PhD defense.