Croft's Question

[tomas]

I posed Bill Croft's question to my colleague Larry Barsalou and below
is the very informative answer I got.

Mike Tomasello
===============


Mike (and Bill),

This issue has been at the heart of the exemplar-prototype debate in the

categorization literature.

Whereas prototypes require a lot of computation during learning (to
abstract prototypes), they require little computation during transer
(matching to a single prototype for a category).  The result is
parsimonious storage (a single prototype).

In contrast, exemplar models have very simple computation during
learning
(the simple recording of an exemplar), but this results in complex
storage
(many exemplars for a category) and complex transfer computations
(matching
the entity to be categorized to a subset or all exemplars).

Surprisingly, perhaps, the evidence overwhelmingly supports exemplar
models.  This suggests that the human cognitive system is not very good
at
abstraction, so it opts for simple learning computations (much work on
concept learning further indicates how bad people are at extracting
rules).
This further suggests that human storage and retrieval are powerful,
given
the human cognitive system seems capable of storing much detailed
information and retrieving it during categorization (as well as matching
it
to the item to be categorized).

I'm embarrassed to say that the paper that probably does the best job of

discussing these tradeoffs is a paper of my own:

Barsalou, L.W., & Hale, C.R. (1993). Components of conceptual
representation: From feature lists to recursive frames. In I. Van
Mechelen,
J. Hampton, R. Michalski, & P. Theuns (Eds.), Categories and concepts:
Theoretical views and inductive data analysis (97-144). San Diego, CA:
Academic Press.

If you or Bill would like a copy, I'd be glad to send it to you.  It
goes
into considerable detail about exemplar, prototype, and connectionist
models on these issues, specifically discussing the costs of storage vs.

computation for each type of model.

A related debate exists in problem solving.  Originally, everyone
believed
that the human cognitive system pieces together rules or productions to
produce solutions to problems (inexpensive storage of a few widely
applicable rules, and expensive computations of chaining them together
while searching a search space).  Now, few people believe that this
characterizes the bulk of human problem solving.  Instead, people appear
to
store cases (i.e., exemplars) and do case-based reasoning (much like
exemplar-based categorization).  The people who have done the most work
on
this are Brian Ross (Psych), Keith Holyoak (Psych), Janet Kolodner (AI),

Kris Hammond (AI).  I'd be glad to send references if you like, but
their
papers are widely available.

Finally, this tradeoff between computation and storage is manifest in
many
current theories of skill.  Essentially, novices are viewed as having
stored few cases, and so have to compute, whereas experts are viewed as
having stored many cases, and so don't have to compute (they just
retrieve).  This goes back to Chase and Simon's work on chunking and
chess,
and it can be found in the modern theories of John Anderson (ACT*),
Gordon
Logan (exemplar-based skill model), and Alan Newell (SOAR).  Each
includes
two ways of producing a behavior--computation vs. retrieval--and assumes

that novices mostly compute but increasinly retrieve as they become
expert.
Again, I'd be happy to send references if you have trouble locating
them.

As you can see, the storage/computation distinction has been central in
psychology for a long time, and it consistently tends to suggest that
the
human cognitive system capitalizes tremendously on complex storage.  I
suspect that the distinction is relevant elsewhere as well.

I hope that this is helpful.  If you have any questions or want to
discuss
anything further, please let me know.

L

P.S.  Mike, I don't have Bill's email address, so if you want him to see

this message, please forward it to him.  Thanks.