perm filename ANLGY.TLK[RDG,DBL] blob
sn#707178 filedate 1983-04-14 generic text, type C, neo UTF8
COMMENT ⊗ VALID 00017 PAGES
C REC PAGE DESCRIPTION
C00001 00001
C00003 00002 Talk ... given to DBL's Learning Seminar, 14-April 1982
C00004 00003 "In terms of level & type of material presented,
C00007 00004 Four subtopics -
C00008 00005 *** Overview ***
C00011 00006 *** [1] What is an analogy? ***
C00014 00007 Does the pair of object alone determine the analogy?
C00016 00008 Two intuitive notions:
C00017 00009 ----- ***** ----- ***** -----
C00020 00010 NB: Relation may not be (explicit) in representation
C00022 00011 We'll use common theory, as helps guide analogy
C00024 00012 Open Questions:
C00026 00013 ... We now know what an analogy is -- that derive "f".
C00028 00014 [3] Reformulation
C00030 00015 [4] Loose Ends
C00033 00016 Other stuff:
C00035 00017 [[ My task ]]
C00044 ENDMK
C⊗;
Talk ... given to DBL's Learning Seminar, 14-April 1982
On Analogy, and things like that
*******
I. Intro
"In terms of level & type of material presented,
this talk, on analogy,
will be like Tom Dietterich's presentation last week on Inductive Learning."
As you recall, Tom presented a coherent framework within which one could
categorize various forms of induction.
I will try to convey a similarly coherent model of analogy.
Information relayed at about the same level.
Fairly formal..., if coarse
Why it useful, & include a rough description of history & other work
(over?) emphasis on speaker's interest/work/prejudices
A relevant question is how (and why) you could have deduced that;
... and not ...
Topic, Speaker, Date/Room, Proportion of time for SubPart
Cohesiveness
amount of preparation I (the speaker) expect of the audience ...
/// coincedental:
same length of time, day of week, institutional affiliation of speaker, ...
///
This deduction is an ANALOGICAL INFERENCE --
On hearing that two things are similar,
you fairly accurately derived the nature of this similarity.
This type of analogical process is topic of this presentation.
----
Also, I want this to be as informal as Tom's
Slides reflect my cowardise more than formality.
- ask questions, make comments
(much for me to learn)
[Note this is a 2nd analogy between this same pair of objects]
Four subtopics -
[1] What is an analogy?
[[Propose a definition to toss around.]]
[[this is bulk - covers/motivates much of the rest]]
[2] What is a good analogy?
[3] Reformulation
[4] Loose Ends
for (less in-)completeness
- grab bag of things skipped or glossed over initially
- my work
---
Not "how to find an analogy", but what it is, being sought.
*** Overview ***
Sketch What it is, and why important,
both in genl, & wrt Learning.
Already sketched an example --
* An analogy involves two objects which are somehow similar.
- describing the analogy means telling the particular way
* [Webster]
Why relevant:
In general - much of cognitive processes based on it -
All familar w/Linguistic - fast communication
Conscious metaphor
"Bill is a pig."
Lakoff
[one structure mapped onto other...]
"He was down in the dumps" vs
"He had his head up in the clouds"
... more generally,
Gen'l Reasoning as well
- patient#1 like that patient#2 [so related treatments]
mutatis mutandis
Simon: Expertise = 50-100K friends
have to be recognized and applied
Carbonell: gen'l problem solving tool
recognition, (e.g. face, similar program), ...
understanding = fitting into familiar framework [= learning]
wrt learning:
[See Psych literature on Learning by Analogy]
Heart of learning -
Accomodation - recognizing new situation as like old
[drawing/exploiting analogy]
Assimilation - incorporating this situation; & indexing
[storing analogy]
[adaptation? - = reformulation?]
---
Why Machine use of analogy?
Theme of seminar (AI in general):
People use it [Ubiquitous], so should machines
Adequate for 1 intelligent processing system, why not another?
[and in particular, in computer interface w/people]
*** [1] What is an analogy? ***
Focus:
What does it mean to "understand an analogical event"?
Problem statement
What is it, *really*?
Just analogues? [# args]
Mapping / Common Theory
Not mechanism, just description
Specific problem is the
Use of Analogy for Learning.
Analogical event:
situation characterized by S telling H a new fact about A,
using the form "A is like B for reason R".
E.g. "talk like Tom's, in terms of contents"
=> You realize that "Russ will present a broad description (model) of analogy".
note: "new"
-- Analogy for purpose of learning.
-- helps specify what type of analogy being communicated
-- not major limitation
S and H might be same person -- remembering
or Expert and KB -- KA
[Yes, many forms of analogy; and many other tasks]
(See Loose Ends)
----
Address few issues:
Note "reason R" clause? Needed?
Does the pair of object alone determine the analogy?
NO!
- needs Context [a third argument]
I gave 2 ways that "This talk like Tom's":
-- coverage of material, & (meta) style of presentation
"Doug used the black board, but this talk like Tom's"
-- i.e. using "slides medium", not black board
"We'd discussed moving this class, but instead this talk will be like Tom's"
--i.e. in same room
or "in same language",
or "by an HPP student",
or ...
Here: can just conjoin - to form "single" analogy.
BUT, might make a difference (lead to contradictory conclusions)
Who is first lady of England?
i.e. US:England :: first-lady ?
Mr Thatcher? Lady Diana? president's wife...
... spouse of head of state
... figure head ...
Consider Washington:Lincoln pairing:
Find ? s.t. W:1 :: L:? - is it 19 or 5 or 9?
Why is Context often not mentioned:
Understood implicitly -- to expect this perspective
In general, one wants the BEST match by some pre-defined "obvious" measure
- why bother w/others
[see next section]
Two intuitive notions:
1. Mapping - certain features correspond.
2. Both analogues satisfy same formula.
[Equivalency]
"Professionally, Doug is like Mike"
Doug:Profession = Mike:Profession
Profession(x) = "Professor"
----- ***** ----- ***** -----
[1] Mapping
Induction-Talk
Topic: Induction, wrt Machine Learning
Depth of coverage: Broad, if shallow
Speaker: TGD
Style: Directed Discussion
Medium: Slides, ...
SubTopics: Essense of Induction, ...
[[ history/other work, why useful ]]
Bias: TGD's interest/work
<< = prejudices>>
Where: MJH 301
When: 2:30-5:30PM, Thursday, 7-April
#NewIdeas: <n>
Analogy-Talk
Topic: Analogy, wrt Machine Learning
Depth of coverage: Broad, if shallow
Speaker: RDG
Style: Open Discussion
Medium: Slides, ...
SubTopics: Essense of Analogy, ...
[[ history/other work, why useful ]]
Bias: RDG's interest/work
Where: 200-205
When: 2:30-5:30PM, Thursday, 14-April
#NewIdeas: 0
(i) Simple approach -- see Depth & Style.
n-tuples of features (R H-R feature set)
(ii) If values must match EXACTLY, would miss lots
Consider
Speaker(ITalk) ~ Speaker(ATalk)
Abstraction hierarchy - as n-th year HPP student
LTalk:Emphasis & ATalk:Emphasis are ONLY similar
-- need to abstract out Speaker
Similarly for LTalk:SubTopics & ATalk:SubTopics, wrt (abstraction of) Topic
LTalk:Time & ATalk:Time -- ignore date field.
So values must be similar.
(iii) Could be stated as equality of values of other (more abstract) slots.
(iv) Relations useful, for perspecuity:
avoids issue of CURRY-ing...
UsefulPreparationFor( x, y, z )
y = Speaker(x)
z = Thesis(x)
PertainsTo( Topic(x), ResearchInterest( Speaker(x)) )
#Ideas: to whom, and what -- for each person
(v) arbitrary formula -- other connectives (in part, =>), existentials
& to interrelate slots, ...
∃ y. Speaker(x, y) & Profession( x, HPP-student)
<<similar to (iv)>>
****
But this is just the common feature (space) approach!
Same formula for both!
NB: Relation may not be (explicit) in representation
Language of talk may be defined as native-language-of(Speaker), ...
Affiliation-of-Speaker...
** so may need to define new values of existing slot, or new slot
[rel'n in general, of course]
----- ***** ----- ***** -----
2. Both analogues satisfy same formula
[See above -- just take conjunction of terms]
f(x) == Depth-of-coverage(x) = "Shallow and Broad" &
Style(x) = "Open Discussion"
----- ***** ----- ***** -----
Equivalency obvious.
I.e. same formula -- with same symbols, to establish connection.
Similarity = equality, in some (perhaps higher) sense
May require reformulation, to see rep'n...
---
Will later [4] describe uses of each --
one better for some applications...
We'll use common theory, as helps guide analogy
(i.e. useful for given analogues, and type of connection, find analogy --
as can store abstractions like Group, or Perspective-as-Human)
----
For learning,
consider theory of hearer, H.
Hearing
Analogous(A B R) R is reason
Find formula f s.t.
Expand(R f) maybe A B as well
where
Th |= f(B)
~ Th |= f(A)
~ Th |= ~f(A)
----------------------
form Th' s.t.
Th' |= Th &
Th' |= f(A)
****
R = "speaker is (HPP) Student"
f(x) = Medium(x, Slides) & Preparation(x TooMuch DBL) &
SubTopics( x, TheoreticalEssense( Topic(x) ) )
<<as student less secure than faculty [slide over bb],
and more head-in-the-cloud theoritical [over empirical] >>
****
Claim that "f" is the analogy --
NOT how to find it, but rather, what it is you're trying to find.
Can relate other work:
Carbonell: "R" = structure of derivation/solution,
f is more abstract fleshing
Winston:-- too confused --
"R" is behaviour of characters (in his abstraction language)
f is specification: Villian(x) will kill King(x)...
Open Questions:
Ideas?
form of reason R?
Reason same form as analogy -- a common relation (together w/meta...)
Like perspective, or guide telling what to map over...
Nature of "Expand" connection.
"Expand" in the form of expanding - to be more specific,...
Subjective, based on USER-INPUT heuristics
EG from "Similar Structure" to "Related SubTopics" and "Level of Detail".
or from "Style" to "Formality Index", ...
Limiting cases: when it is conjectural vs deduction...
Mechanizable form -- based on syntactic...
----
Contributions:
*** obvious ***
(1) 3rd arg - perspective, context, guide for what to carry over.
(2) Unification of Map & Commonality
[modulo reformulation]
*** decomposition of problem ***
(3) Only R in Expand - not A,B
(4) Seperation of Novelty ++ later over conditions on A,B.
... We now know what an analogy is -- that derive "f".
[2] What is a good analogy?
-- What is it that makes an analogy good?
[useful, applicable, relevant]
-- Is analogy#1 better than analogy#2?
Need:
As shown above,
There may be >1 analogy between 2 given analogues.
Much work - most of AI works
See Gentner, Carbonell, (other work: Invariance Hierarchy) ... etc
Applications:
Evans: Educing [many R s.t. R(A,B) -- want one s.t. R(C,i)]
Winston: most similar play - nature/behavior of characters
AP: best similar program - functionality
...
In our model,
there are many possible "f"s, even given A B & R.
Clearly: task/goal dependent
[often implicit, "obvious"]
E.g. "Best analogy is one which solves the problem."
I see this is being quite subjective -- to be guided by heuristics...
[All syntactic, trying to emulate depth of association]
Intution:
"Deeper" (more semantic) is better
How easy it is to find (or measure) is clearly dependent on the rep'n
[3] Reformulation
------
Problem: Analogy is a relation between Analogues,
NOT their representations!
(I.e. not dependent on formulation for acceptance.)
But, when mechanized,
(to implement)
facts are input in a rep'n.
As w/any problem, the soln is obvious in correct rep'n.
"Ewe:Lamb :: cow: ?" -- sheep, not bovine
In ideal rep'n,
"when nature is cuts at its joints"
abstraction is obvious;
just need simple (parameter) instantiation... explaining why
Any analogy is obvious in retrospect.
If wrong, need to change. (to find best abstraction)
saw some simple cases above - new slots from old variety
Consider ? circle/sphere -- clearly better in polar coordinates over cartesian!
Grab bag of techniques --
new relations from old,
Abstracting/focussing,
coordinate change
[see Amarel]
Major lesson here:
Analogy = Reformulation + (Simple) Instantiation
Especially for one person (program) to understand another.
[4] Loose Ends
My focus fairly sharp.
Mention other aspects:
[Why: all analogy relevant to learning, in that learning
involves recognizing and applying similar situation]
3 Types of analogy
- similar,
- proportional (educing R from A & B -- another degree of freedom)
... actually continuum...
simile
- familial
Family resemblence, Quine's ostention, Wittgenstein "game"
[other sense of analogical reasoning - by prototype, not primitive]
Types of Analogy tasks
- find (best) analogy from analogues (w/direction)
for some task - eg A:B :: C:?, or ...
- find (generate) analogue from other analogue
AP, precedent in Legal reasoning
- compare/rank analogies
- use analogy for guidance
Kling, Carbonell
- use the analogy to deduce/conjecture new facts
about one of the analogues
[quickly describe that analogue]
- use the analogy to form an extended analogy
(and then used as above)
Conclusivenss
Deduction
Conjectural
(based on reliability of that Extend operator)
* consider R like T *
Level - model (broad) vs instance
Model to Model: Electricity is like Water Flow.
Instance to Instance: This talk like Tom's.
Instance to Model: <These have the basic flavor of induction.>
or instanceS -- to classify, find category, ...
Model to Instance: <These seems simple instantiation.>
Inheritance from prototype
Other stuff:
Literature search
My work
--------
Limitations of mappings
(1) if only slots, too limited.
(2) (either way) - how to get the analogy from analogues and "context"
no way to store the commonality
(which could be used for guidance -- to suggest needed reformulation)
Limitation of Common Theory
(1) hard to express differences (likewise: "don't-cares")
(2) hard to express correspondences
-------
*** Philosophy: ***
ala Polya (see also Dershowitz, JSBrown, ...)
Two analogues are siblings in some abstraction space --
with a common abstraction, but different instantiations.
Q: ∃? a reality to an Analogy, or just linguistic convenience?
*** Psychology: ***
Tveseky - similarity studies
Anderson et al, 1978 - ACT
[[ My task ]]
My task:
Focus on use of given analogy.
First finding the connection --
involves first finding the apt representation
& then match.
Then exploit -- trying to transfer other facts over,
to flesh out new analogue.
(based on heuristics).
Why?
Clearly learning - updating KB.
Most learning is fairly constrained...
∃ goal of solving some task to guide search.
[Meta] testable, sorta
*****
Examples of Learning - Understanding something new
Shallow - ala copy&edit
VM like BM
Deeper - pedagogic tool
Given FM, learn EC
- to "understand/predict", and diagnose
- then Unix (other pipes system)
Text editor from <typewriter>, <secr'y>
[Will later discuss differences - relating to reformulation]