Deep Analogical Inference as the Origin of Hypotheses | Mark Blokpoel, Todd Wareham, Pim Haselager, Ivan Toni, Iris van Rooij; Journal of Problem Solving, 2018 vol. 11, 1-...
Where do candidate hypotheses for inference to the best explanation come from?
The origin of hypotheses, otherwise known as abduction proper, is hallmarked by seven properties: (1) isotropy, (2) open-endedness, (3) novelty, (4) groundedness, (5) sensibility, (6) psychological realism, and (7) computational tractability. In this paper we provide a computational-level theory of abduction proper that unifies the first six of these properties and lays the groundwork for the seventh property of computational tractability. We conjecture that abduction proper is best seen as a process of deep analogical inference. (1)
So notice that this concerns abduction proper whereas IBE usually assumes the hypotheses are already available.
The main effort of this paper is to unify work done on these seves aspects into one theory.
Also notice that the proposal is on the computational level.
Properties
Isotropy: Well known property advocated by Fodor (1983). "Any knowledge that a person has can potentially be relevant for making the abductive inference" (2).
Fodor really means any knowledge. An inference may be affected by knowledge that seems unrelated at first sight.
Open-endedness: the set of candidate hypotheses may contain every hypothesis that someone can possibly generate. Cf. Goodman, 1983. This could mean an infinite amount of hypotheses.
Novelty: the hypotheses do not have to be generted before, but can be new. Again, see Fodor and Goodman in 1983. E.g. one may use a gesture to indicate a particular meaning for the first time; to correctly infer this, one cannot assume a pre-existing hypotheses about this novel use of the gesture.
Groundedness: "A candidate hypothesis must have a well-defined relation to a representation of the observation that is to be explained" (2). See Lakoff & Johnson, 2003.
Types of observations:
- Perceptual observations
- Internal observations (e.g. a scientific conclusion)
Minimally the theory should handle "hypotheses being grounded into perceptual representations" (2)
"In the case of perceptual observations, this property can be construed as classic groundedness" (2) See Barsalou, 1999.
Q: When discussing groundedness, two types of observations that could be explained are distinguished: perceptual observations, and internal observations. It is then stated that the theory should minimally handle "hypotheses being grounded into perceptual representations" (2)
I'm not sure if that means that we assume from now on that these "internal observations" are a special case of perceptual observations, or if we are leaving aside the issue for now because internal observations cause extra complexity w.r.t. groundedness.
Why minimally grounded in perceptual representations? Does that mean that the assumption is made in the paper that "internal observations" are also representable by perceptual representations? (2)
Compare remark on page 6: "Note that each hypothesis contains a well-defined relation between the core and the evidence e (see also Figures 1 and 2). Hence, the theory is grounded even in the classical sense since the evidence can (but need not) be perceptual.
Sensibility: the best explanation does not necesarily have a high likelihood. Van Fraassen, 1985: "the best of a bad lot". In other words, the set of candidate hypotheses should contain "only sensible candidate hypotheses" (2).
Q: This requirement seems a bit strict. Would it be sufficient to say that the set should contain sufficient sensible candidate hypotheses?
What is sensible?
- Must be grounded (see previous point)
- Must actually explain the observation
Psychological realism: this matter is decided based on empirical research of the computational account. But the point is simple: if empirical research shows constraints on the inferences that can be made, those constraints should be taken into account.
Computational tractability: Human cognition is resource-bounded, so this has to be taken into account. There is a clear tension with open-endedness and isotropy.
We propose this unification can be achieved by viewing the origin of hypotheses as a process of deep analogical inference. Whereas a single analogical inference finds one structural relation between two representations, deep analogical inference allows many consecutive and branching analogical inferences that lead to sets of candidate hypotheses. (2)
Abductive inference
Abductive inference is a combination of abduction proper and IBE.
This follows the terminology by Haselager 1997 and Lipton 1991.
This paper formalizes abductive inference as an input-output mapping on the computational level. Even though algorithmically the two steps of may be intertwined, they can be described separately on a functional level.
Q: on p.3 various notions of "best" are discussed. Curiously, Lipton is cited for "most probable" and "most likely", which is surprising to me because I've come to known Lipton as arguing against this interpretation, instead arguing for an interpretation of "best" as the most "lovely" explanation.
Approach:
This paper uses the structure-mapping theory (SMT) of analogy to specify the exact "relationship between a set of candidate hypotheses and the evidence and knowledge" (3)
Analogical origin of hypotheses
Analogical reasoning has been identified as important in "domains that require creative leaps" (3)
Mentioned examples (look at the original paper for many references). These are quoted:
- "language learning and understanding"
- "concept learning"
- "problem solving"
- "similarity judgement"
- "scientific explanation"
- "perception"
- "generalization"
Why SMT?
SMT is chosen as the account of analogical reasoning because it is one of the few accounts on the computational level. (3) The theory can be used to provide constraints for algorithmic level accounts.
SMT also already covers novelty, sensibility, and psychological realism. (3)
- An analogical mapping may be across domains, accounting for novel projections
- "If, in an analogical inference, one of the two representations is that of an observation, then the resulting hypothesis can be used to explain that observation by analogy"
- There is empirical support for SMT in cognition modelling. Presumably then, hypotheses generated using SMT will be psychologically plausible.
To also explain isotropy, open-endedness, groundedness, and computational tractability, an extension toSMT will be made using analogical matching and deep analogical inference.
Approach:
First, all candidate hypotheses generated by deep analogical inference are potentially grounded, because deep analogical inference guarantees a relationship between the observation and the can- didate hypothesis, and the representation of the observation can be perceptual in nature. Second, the set of candidate hypotheses is also open-ended, as the characterization includes all possible hypotheses a person can in principle generate. Third, isotropy is guaranteed because all knowledge representations available to a person serve as a potential link in the chain of deep analogical inference. Finally, by formalizing the theory at the computa- tional level, we will lay the groundwork for investigating under which conditions it is computationally tractable (cf. van Rooij, 2008; van Rooij, Evans, Müller, Gedge, & Wareham, 2008). (4)
What are the current accounts of abduction proper?
- lambda-calcululs framework by Goodman et al. 2008
- hierarchical Bayesian models: generate hypotheses using a grammar
- "bridge symbolic representations with probabilistic computations" (4)
- SMT
The first two accounts could potentially account for the seven properties, but it is currently not yet clear how they do.
Structure-Mapping theory
- You have a base and a target between which you can find an analogical match.
- Matches are determined using structural overlap between relational representations
Given a match, you can project inference from base to target. (4)
Relational representation
Representations are made up of objects, attributes, functions (1 arg) and relations (2+ args).
Analogical matching
Correspondences between entities in the representations. (I think "entity" is defined broadly here as objects, attriutes etc.)
Rules:
- Objects, attributes and functions can only match entities of the same type
- Relational entitities can only correspond to entities of the same type and with the same label
- What does that mean?
-
Matches have to be structurally constistent:
-
"1:1 correspondence: Each entity that is part of the match can only be part of one correspondence." (5)
- "Parallel connectivity: If an entity is part of the match, then all its arguments should also be part of the match." (5)
Example on page 5:
A kid called James believes that "kicking a can is good pretend play for playing soccer" (5)
- kicked(isGirl(Mary), hasColor(isSphere(Ball), Red)
- kicked(isBoy(James), hasColor(isCylyndrical(Can), Silver))
Q: The paper says all arguments also corresponds, but isn't there a difference between isGirl and isBoy? Is it assumed here they are each others negation, logically? We see that kicked corresponds. That's enough? But all arguments should also be part of the match, right? Maybe I misunderstood what correspondence means here.
This inference remains sound even in another domain "because structural overlap is guaranteed and because labels of objects, attributes, and functions can be ignored" (5)
The paper also says that labels can be ignored, except for relations. But both kicked and hasColor are relations and use different labels. What's up with that?
Maybe I misunderstand the technical meaning of 'labels' here.
The quality of an analogical match is determined by its systematicity. A match has high systematicity "the more the analogical match is interconnected and the more deeply nested substructures it contains" (5) (bunch of references here)
Inference projection
Main constraint: "the projected part of the base has to connect to at least one attribute, function, or relation in the overlapping structure." (5)
Other possible constraints (see paper for references):
- Goal relevance
- Adaptability
- support and/or extrapolation
The main benefit of projections is that they can work across domains.
Variable instantiation
If there is a match between two representations, we can also transfer knowledge by instantiating an object in the target representation with an object from the base representation.
To summarize, projection transfers structures, and variable instantiation transfers objects.
Candidate hypotheses
Hypotheses quintuple < b,t,m,p,i >
- b: base reprepresentation
- t: target reprepresentation
- m: analogical match
- p: projection function for structures from b to t
- i: instantiation function for objects from b to t
Note that the resulting structure is only a candidate hypothesis. The result of the analogical information transfer is not guaranteed to be correct!
Abduction proper by deep analogical inference
Recursive analogical matching and inference projection.
"All candidate hypotheses in the output are based on the evidence e, which guarantees that all candidate hypotheses are grounded." (6)
Q: The grounding relation is still a bit implicit for me. And how does the evidence come into play in the formulas? Because the base representation describes an observation right?
The output is "all possible candidate hypotheses for e relative to a core k" (6)
The concept of core is newly introduced. What does it mean?
ABDUCTION PROPER:
Isotropy is met by considering the union of all hypotheses for all cores k in K, where K is a set of relational representations of knowledge.
So I guess a core is one piece of knowledge, expressed by a single structure.
For this we need a set of hypotheses with respect to a single core k.
This is returned by ANALOGICAL CANDIDATE HYPOTHESES(e,K,k), which outputs "a set of cancidate hypotheses for e relative to a core k" (6).
These < e', k', m, p, i > "are based on every analogical match m, project p, and instantiation i" that can be found between all possible representations of the evidence e and all possible representations of the core k" (6).
REVIEW.
This strategy is taken so that no relevant candidate hypotheses are excluded. The debate is out on methods for determining how much is projected and instantiated in the process.
Q: How do you determine all possible representations for given evidence e?
DEEP ANALOGICAL INFERENCE
Recursively applies ANALOGICAL AUGMENTATION as often as possible.
AA returns a projection with the highest systematicity (i.e. "the biggest possible structure from k to a")
Result: set of all possible representations of a (so that answers the above question).
This also contributes to isotropy.
The third and final part of the theory that contributes to its isotropy is that deep analogical inference returns sets of representations that are based on all possible sequences of analogical inferences (i.e., match and projection) with all knowledge. (7)
Which also contributes to open-endedness.
Impression of the flow
DEEP ANALOGICAL INFERENCE to find all possible representations of a (w.r.t. k?)
- We have a relational representation a and are trying to find all possible relations of a
- a analogically matches knowledge k
- and can then be augmented via analogical projection into b
- repeat on b until the result does not match any knowledge k in K
Insight:
"if a is a perceptual representation, then each representation built on top of a is grounded in (i.e., has a well-defined relationship with) a by virtue of the transitive analogical relation"
These are also called "reconceptualisations" in the paper.
ANALOGICAL ABDUCTION PROPER:
Given all representations in the space of all evidence representations; and the space for some core k, you can find matches and projections.
Fodor's frame problem
The complexity analysis of this theory is essentially unsolved. There is a clear paradox between open-endedness and isotropy on the one hand, and the fact that humans can do quick inferences with resource-bounded cognition.
The authors argue this is not a reason to reject this theory. Rather:
The fact that many theories of abductive inference (including analog- ical abduction proper) are computationally intractable (Bylander, Allemang, Tanner, & Josephson, 1991; Nordh & Zanuttini, 2005) can be seen as a sign that cognitive science is currently unable to solve Fodor’s frame problem: How can abductive inference be isotropic, yet computationally explained (Fodor, 2000)? Where Fodor was pessimistic about the chances of computational cognitive science fully solving this problem, we are not and propose a way forward in the section below. (9)
This move forward is parametrized complexity analysis.
The reason for our optimism is based on the fact that computational intractability is not a property of the size of the search space (even if it is infinite), but of the ability to search that space efficiently. (9)
So there is an option ways can be found to leverage the structure of the search space, however big it is.
Q: 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 amounts 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.
Appendix: case study
Target: innovative communicative signal in a communication game (i.e. a novel signal with a novel meaning that requires novel hypotheses).
The capacity to generate and understand novel signals is the topic of experimental semiotics. The main methodological challenge is to study novel signals in isolation because they are often interspersed with conventional signals.
See the Tacit Communication Game .
Note that identical signals can have many different representations and meanings; and as a consequence, different behavioral replies. The same behavioral response of the receiver can also be the result of different signal meanings (16).
E.g. a repetitive wiggle between A and B could mean "move from A to B". Or the amount of repetitions could indicate something about the direction to use, or the number of rotations to perform (e.g. a circle shaped token cannot indicate rotations by itself!).
The wiggle is a communicative innovation generated by senders to indicate location and orientation of a receiver’s token, when the sender’s token has less rotational freedom than that of the receiver. For example, in trials where the sender has to use a circle (which cannot show rotations) to communicate the orientation of a triangle (which has four orientations), often—but not always—senders adopt a wiggle signal. (17)
The wiggle signal consists of the communicator pausing her token at the receiver’s target location to convey that his token should be positioned there. Then, unable to orient her own circular token, the sender uses repetitive movements along an axis to signal the orientation of the receiver’s triangle (see Figure A3). This signal, however, can mean various things within the context of the game. For instance, it can mean the pointing direction of a triangle (“wiggle apex”), or the number of times a receiver needs to perform a “rotate” action (“wiggle rotate”), or even the opposite of the pointing direction depending on the number of repetitions (“wiggle opposite”; see Table A1).