Greg Detre
Monday, October 28, 2002
1.
For
each paper, summarize the main ideas and your reactions to them.
2.
Where
does Regier & Carlson�s model of spatial semantics fit into Gardenfor�s
scheme?
3.
Consider
the following syllogism:
Cheap apartments are rare
Rare things are expensive
---------------------------------------
Cheap apartments are expensive
We can tighten the language and retain the
result:
Cheap apartments (in Boston that are desirable)
are rare (things)
Rare things (in Boston that are desirable) are expensive
-----------------------------------------------------------------------------
Cheap apartments (in Boston that are desirable) are expensive
Provide an explanation for why this syllogism
fails to produce a �reasonable� conclusion.
�Intrinsic
representations resemble [i.e. their internal structure is isomorphic
to] what they represent. In contrast, extrinsic representations must be
accompanied by a rule which specifies how the representation is to be
interpreted � such a rule provides the �meaning� of the representation�.
If
Gardenfors is happy to say that the conceptual level is grounded, then I don�t
see how we have a problem grounding the symbolic.
I really
like the intrinsic/extrinsic representation distinction � presumably intrinsic
representations are isomorphic, right???
it�s not obvious to me that though Gardenfors� way of describing conceptual spaces makes them seem very attractive, that they aren�t in fact continuous with a symbolic representation, and indeed subject to many of the same problems
aren�t they simply (intrinsic) symbolic representations???
is the idea of an intrinsic symbolic representation an oxymoron???
what happens if the space is extremely high-dimensional???
e.g. smell�???
are you supposed to only choose/capture the important quality-dimensions???
if so, then isn�t their �meaning� derived from ours in the same way that symbolic meaning is???
really important: how do you figure out what your quality dimensions are, i.e. how do you learn them???
are they any more dynamic than symbolic systems???
relatedly, how are they grounded???
is the conceptual level supposed to be continuous with both symbolic and connectionist???
if so, that would explain some things, e.g. that you could use connectionist self-organising maps to find the principle components, for example, and translate these into your dimensions, and then build up from there into the symbolic level�
but if that�s what you�re trying to do, what does the symbolic add??? do you do computations on/at the conceptual level??? does he imagine that this would push the symbolic level into just a few select, narrow domains???
ok, he talks about the conceptual �level� and �perspective� etc. rather than the conceptual space (as in �activation space�), but at the same time he talks of it being of lower dimensionality etc. � how exactly does that reduction in dimensionality happen, if it�s not an explicit computation, and given that you can�t always be sure that a lower dimensionality will emerge???
what does Quine mean by �alternation� as a notion of logic??? � he means �or�
Gardenfors
has read Harnad on the symbol grounding problem � how does his account
differ???
well, the main difference is that he�s proposing a different
intermediary level � rather than iconic representations, he�s suggesting the
conceptual level, which is a kind of low-dimensional vector space
is it a lossless reduction from connectionist to conceptual and
conceptual to symbolic???
I doubt it, when you consider the enormous loss between the
connectionist and symbolic representations
I just
don�t think he�s clear enough about the mechanics of the conceptual level. I
don�t fully understand how computations would be performed. I can�t tell
whether it�s actually a level, or simply an abstraction that we can choose to
make when viewing the overall activity/behaviour of a neural network. He says
somewhere that he doesn�t want to make an ontological statement about the
existence of the three levels, and that it emerges out of the connectionist
representation, which I suppose settles the discussion, but it�s not even clear
what tools you could use to see things at the conceptual level. Do you do a
statistical analysis of the activity of the NN???
� he does go on afterwards to talk in terms of vector spaces and vector calculations, but I�m still not sure how he calculates the vector space that the conceptual level operates in in the first place�
can we be a
bit more precise about why it�s �unnatural to view such [connectionist]
calculations as examples of rule following�???
holistic vs
distributed???
what can I
take away as positive/constructive from Gardenfors� article???
above -
(though they have good reason to view this as just one example of a� , - left, right, down etc.
did they
compare the acceptability ratings for above for the positions below the
half-way line, as a way of seeing which positions we think are most opposite of
above???
did they even take experimental results for those below positions???
yes, I think they did�???
�microworld
whose sparseness is disguised by a paper-thin fa�e of real-world terms�
Hofstadter
charges �that Holyoak and Thagard failed to recognize the gulf between human
perception of a situation which has no hard and fast distinctions and the view
of a situation as a set of predicate calculus formulas full of distinct
objects, attributes, and relations� (quote from the summary of the review,
not the review)
�With their new analogy-maker, LISA, Hummel and
Holyoak use a connectionist hybrid network to incorporate semantics into their
analogical model � The major improvement over ACME is that predicate elements
are now connected to semantic units which supply a rudimentary meaning to the
proposition. A pair of analogous propositions will share many semantic units,
and infrences can be made about one proposition from the semantic relationship
with its analog�
�Hummel and Holyoak claim that LISA solves the
problem of maintaining structure in a distributed representation, but they fail
to address the major criticism posed by Hofstadter in his critique of the earlier
model. LISA still uses formalized predicates, and still focuses on
correspondences between propositional elements�
Analogy is
(1) similarity in which the same relations hold between different domains or
systems; (2) inference that if two things agree in certain respects then they
probably agree in others. These two senses are related, as discussed below. �
Conflating
�tiny, inert predicate calculus cores� with �the original full-blown
situations�
�frozen,
caricatured
definitions
of analogy � is isomorphism key???
would it be
fair to say that the conceptual level is what Hofstadter has in mind for
analogy-making???
either way, it should be pretty clear that the level is key
can we make analogies at multiple/every level???
are we seeing analogies between the gists, the facts, the structures,
what??? all of those things???
cognitive
linguistics???
i'm trying
to think of it in terms of supply and demand terms, at least initially, to
tease apart what feels like an ambiguity in the word 'cheap'.
if we take
the first syllogism, we can read the first sentence as saying one of the
following:
a)
there
are very few nice (i.e. for which there exists high demand) apartments
available below the market price
b)
there
are very few apartments at all (i.e. nice or otherwise) below a certain
arbitrary price
c)
there
are very few cheap-and-nasty apartments
if we then
look at the second sentence of the first syllogism, we can see it as saying:
a)
things
that are desirable but in low supply have a high market price
the
conclusion of the first syllogism only makes sense if we read it as:
�
if we look
at the first sentence of the second syllogism:
a) it supports reading (a) of the first sentence of the first syllogism
if we look at the second sentence of the second syllogism:
2nd vs 1st
syllogism
unevenly distributed � in Boston
�that are desirable�
possible
ambiguities
expensive = propensity to rise in price
rare = small subset of
cheap = below what it�s worth vs cheap-and-nasty
misc ideas
they�re rare because they�re not expensive
transitory
reject the second premise
cheapness as an extrinsic property that can be attached/detached
cheapness itself is rare in desirable things
why is cheapness a special case?
cheap + expensive are related/opposites
they�re a function of rareness + desirability�???
new syllogism:
desirable things are generally expensive
\ cheap desirable things are rare
why are we looking at the syllogism in the first place???
is it just an exercise to get us to think about meaning, and how different terms mean different things when we are thinking within different domains???
is it because we are going to read something relevant soon???
I don�t think it�s that
does that mean that I should see parallels with the articles on conceptual space, analogy and spatial terms???
what are the links between them alone anyway???
they�re
all about getting to the gist
does he want us to do an economic analysis of the terms and show the ambiguity in that way???
why have two syllogisms been provided??? why not just the second??? what explanation for the ambiguity might we come up with if we were only looking at the first that is precluded by the second???
are we going to discuss this in class???