COGSCI 401--FINAL EXAM UPDATED ACTUAL Questions and
CORRECT Answers
[S13, Analogies] What is an analogy? An analogy is mapping the structure of one situation onto another, usually across
domains.
This process involves:
--Recognizing similarities in relationships, not just objects.
--Applying known relational structure to a novel context.
**not rare, human reasoning makes analogies to previously experienced
situations.
[S13, Analogies] Dierdre Gentner's SME (Structure A symbolic model of how people make analogies, using structured
Mapping Engine) representations and explicit rules to match situations.
[S13, Analogies] SME: Knowledge Representations --The mental models of two situations in comparison.
--Symbolically encodes things, actions and relations.
EX: Sun (base) and Atom (target)
Base: ATTRACT(SUN, EARTH), ORBITS(EARTH, SUN)
Target: ATTRACT(NUCLEUS, ELECTRON), ORBITS(ELECTRON, NUCLEUS)
, [S13, Analogies] SME: Mapping Rules OCCURS AFTER KNOWLEDGE REPRESENTATIONS, SME uses mapping rules to
see what matches between A and B
Aligns relational structures: "If two things play the same role in the same kind of
relationship, they should be mapped to each other."
In our example:
ATTRACT(SUN, EARTH) → ATTRACT(NUCLEUS, ELECTRON)
Both have a central object attracting a smaller one.
ORBITS(EARTH, SUN) → ORBITS(ELECTRON, NUCLEUS)
Same kind of relational structure.
It maps:
SUN → NUCLEUS
EARTH → ELECTRON
[S13, Analogies] SME: The Bitter Lesson (1) SME uses hand-coded symbolic representations of knowledge.
(2) SME uses hand-crafted mapping rules
ASK: (1) Does SME learn anything from data? No. (2) Is SME's performance
dependent on computation and scale? No.
In Bitter Lesson Approach (LLMs, deep learning): learned from massive datasets,
flexible learned representations, gets better with more data and compute, scales
to real-world complexity.
**SME is exactly the kind of hand-crafted system that the Bitter Lesson deters us
from using. SME is helpful for theory and small examples, but LLMs now are better
than SME's at analogy making.
Rich Sutton's Bitter Lesson is that general methods that leverage computation and scale (like self-supervised
learning) eventually outperform hand-coded knowledge and human-designed
strategies.
BASICALLY: Stop hardcoding knowledge. Let the machine learn it.
[S13, Analogies] Why are LLMs so good at analogies? Because LLMs use word embedding spaces to naturally support analogical
reasoning.
WORD EMBEDDING SPACES: multidimensional space where words are
represented as vectors, to characterize features that a word might have on a
scale.
[S13, Analogy] What is the female analogue of King? King -- Male + Female = Queen
What about the royalness?
King -- Royal + Commoner = Male Commoner
[S14, Theory of Mind] What is Theory of Mind? ToM is the ability to understand that others can hold beliefs that are (1) false and
(2) different from one's own.
CORRECT Answers
[S13, Analogies] What is an analogy? An analogy is mapping the structure of one situation onto another, usually across
domains.
This process involves:
--Recognizing similarities in relationships, not just objects.
--Applying known relational structure to a novel context.
**not rare, human reasoning makes analogies to previously experienced
situations.
[S13, Analogies] Dierdre Gentner's SME (Structure A symbolic model of how people make analogies, using structured
Mapping Engine) representations and explicit rules to match situations.
[S13, Analogies] SME: Knowledge Representations --The mental models of two situations in comparison.
--Symbolically encodes things, actions and relations.
EX: Sun (base) and Atom (target)
Base: ATTRACT(SUN, EARTH), ORBITS(EARTH, SUN)
Target: ATTRACT(NUCLEUS, ELECTRON), ORBITS(ELECTRON, NUCLEUS)
, [S13, Analogies] SME: Mapping Rules OCCURS AFTER KNOWLEDGE REPRESENTATIONS, SME uses mapping rules to
see what matches between A and B
Aligns relational structures: "If two things play the same role in the same kind of
relationship, they should be mapped to each other."
In our example:
ATTRACT(SUN, EARTH) → ATTRACT(NUCLEUS, ELECTRON)
Both have a central object attracting a smaller one.
ORBITS(EARTH, SUN) → ORBITS(ELECTRON, NUCLEUS)
Same kind of relational structure.
It maps:
SUN → NUCLEUS
EARTH → ELECTRON
[S13, Analogies] SME: The Bitter Lesson (1) SME uses hand-coded symbolic representations of knowledge.
(2) SME uses hand-crafted mapping rules
ASK: (1) Does SME learn anything from data? No. (2) Is SME's performance
dependent on computation and scale? No.
In Bitter Lesson Approach (LLMs, deep learning): learned from massive datasets,
flexible learned representations, gets better with more data and compute, scales
to real-world complexity.
**SME is exactly the kind of hand-crafted system that the Bitter Lesson deters us
from using. SME is helpful for theory and small examples, but LLMs now are better
than SME's at analogy making.
Rich Sutton's Bitter Lesson is that general methods that leverage computation and scale (like self-supervised
learning) eventually outperform hand-coded knowledge and human-designed
strategies.
BASICALLY: Stop hardcoding knowledge. Let the machine learn it.
[S13, Analogies] Why are LLMs so good at analogies? Because LLMs use word embedding spaces to naturally support analogical
reasoning.
WORD EMBEDDING SPACES: multidimensional space where words are
represented as vectors, to characterize features that a word might have on a
scale.
[S13, Analogy] What is the female analogue of King? King -- Male + Female = Queen
What about the royalness?
King -- Royal + Commoner = Male Commoner
[S14, Theory of Mind] What is Theory of Mind? ToM is the ability to understand that others can hold beliefs that are (1) false and
(2) different from one's own.
