Greg Detre
Tuesday, October 29, 2002
cheap
apartments problem � came out of OpenMind � employs modus ponens and other
syllogistic machinery
interesting, because there�s lots of things going on, including not
being clear about it
Dennett�s
formulation of the Frame problem
in terms of a robot that has to get a wagon out of the room without
taking the bomb with it
you have to not only know what�s relevant, but not compute on what�s
irrelevant
for symbolic systems, there�s no way of figuring out what�s relevant
just from the structure of the statements
�frame problem� is a phrase from psychology (rather than
Minsky/McCarthy), meaning a frame of reference within which all of the issues
are contained
Dennett doesn�t see this as a purely symbolic problem
probabilistic (e.g. Bayesian) systems still have problems with this
�???
innateness
+ symbol grounding (vs creation) problem
induction � how can you do that without some sort of semantics attached
to the symbol
Deb: is there any meat to the symbol-grounding problem?
what is symbol grounding??? grounded to what???
senses???
Harnad, symbol grounding problem: �How can the semantic interpretation
of a formal symbol system be made intrinsic to the system, rather than just
parasitic on the meanings in our heads? How can the meanings of the meaningless
symbol tokens, manipulated solely on the basis of their (arbitrary) shapes, be
grounded in anything but other meaningless symbols?�
connectionism:
Peter: usually the NNs that work well are the ones that aren�t opaque,
where you�ve done a lot of engineering to pin down a structure that does what
you want to do, and that you can understand to some degree � Gardenfors�
description is a highly idealised description
conceptual
level:
there�s some meaning to each vector
notion of similarity
object/instance as point in low-dimensional quality space, properties as
regions
can you differentiate individual objects/instances in this low-dimensional
space??? can you interpret the properties??? is there enough information??? is
the conceptual level really above the symbolic???
two cups with the same features couldn�t be told apart at the conceptual
level
Hugo: unhappy with the idea of similarity as Cartesian distance
his examples are very clean + deceptive
non-geometric space???
Tom: this sort of geometric representation doesn�t work so well for more
conceptual stimuli, e.g. animal taxonomies, or diseases
for more structured stimuli, you�re going to need a higher-dimensional
space � you can do animal taxonomies, if you look at all of their
distinguishing features
the main thing you need to deal with space is a
notion of betweenness
is the conceptual level simply the symbolic???
Hugo: he says so at the end
Barbara: why can�t you build it in a connectionist way
but we might not understand what�s going on internally in the
connectionist model
Deb: having nameable dimensions doesn�t necessarily mean that it�s
symbolic
Deb: connection to metaphor? does having a lower dimensional space make
analogy-making easier?
Tom: a lot of symbolic reasoning lacks analogue structures, which they
need
surely though real numbers are analogue???
allows you to use geometrical isomorphisms
see them as caricatures of NNs
principle component analysis on the hidden layers
more information at the symbolic level???
when you put a label on a dimension, you impute meaning, you turn it
into a symbol
???
pairwise relations between points as a means of 2D representation???
Hugo: even �above� would be much more complicated than the examples
Gardenfors gives
Deb: no problem with having hundreds of dimensions
but then �conceptual� becomes a misnomer???
what�s the appeal then???
need temporal (and distance) information too
Hugo: ok, what about emotions then?
Peter: the taste example is perhaps misleading because you�re not
necsessarily supposed to find a physical correlate for the quality space
are you supposed to be able to put a name to
the quality dimensions???
summarise
complaints:
there are many high-dimensional examples that wouldn�t fit in low
dimensions
mainly constrained to perceptual spaces
maybe there
should be many things in between the connectionist and symbolic (cf Minsky)
one big
thing that we can take away from Gardenfors� ideas is that you can leverage the intrinsic structure of your representation
this does help with the grounding
so how does
Gardenfors deal with the symbol creation problem???
difference between you can�t yet and won�t ever be able to tell us how
that happens
they all
seem to think that the conceptual is closer to a sophisticated view of
connectionism than to the symbolic
I should have looked at Harnad�s 8 rules of the symbolic � the symbolic
is discrete
you could
put the centre of your attentional exponential fall-off maybe in the centre of
mass � but then it wouldn�t reduce to both the PC and COM models
would you
have �above�, �left� etc. as quality-dimensions???
or would you have the �absolute� spatial dimensions as the
quality-dimensions???
Regier assumes that there�s always an absolute �up�
Regier�s
model of above is not symbolic because it has non-discrete components
the
original Regier work is all connectionist
Deb: are
non-point targets an easy extension?
Peter: no, because of the way he finds attention points
Deb: you can just use the proximal for each, just as you do for the
landmark alone
what about
multiple objects?
changes your notion of COM, depending on whether you count the balls or
their total mass
problem
with the human experiments???
artefact introduced by making our concepts explicit??? could we not make
things a bit more natural, and divert their attention from an examination of
�aboveness�
what if people got into some algorithmic habit?
0-10 vs �1-10
part of the problem is that it�s 2D
sometimes,
you test aboveness by dropping the pen, and seeing if it lands on the table �
depends on notion of down/gravity
of course, then helium balloons would never be above anything
Deb�s talk
tomorrow
you need a really good reason to define a level
�leap of faith� processes � there�s Inference to go from one system to
another
he doesn�t see that in Gardenfors� connectionist to conceptual
grounded/sensorimotor is the first leap of faith
objective view � seeing the world as objects (track,
co-index, same object as before) seems to be the second leap of faith
third leap of faith � Dennett�s intentional stance
(object moving with speed vs determination)
fourth level (Jackendoff) � legal systems in ethics, a
person level
�levels� implies that they sit on top of each other � whereas these are
more like big domains
realms??? stances???
on what basis should we even declare a level?
is it that they self-discretise???
different purposes: unit of analysis � demarcate certain behaviours,
perhaps because of the problems it has to solve
levels of intentionality???
Dennett�s stances:
physical
design � knowing what it�s designed to do let�s you abstract away the
gears and springs (alarm clock)
intentional � let�s you talk in terms of goals, beliefs and rational
planning processes
Deb: is he justified in separating design from
both intentionality (function) from physical? there�s no discontinuity
design as a derived goal (i.e. a subset of intentionality)???
the stances are about (convenient/shortcut)
ways of interacting with things
Dennett: Martian that can predict all the atoms of the universe � it
would be amazed that we have the predictive power that we do
is Deb discussing what�s useful about, or
saying (philosophically) that these are the levels?
Hugo will
post onto the common sense bulletin board
Deb talk
tomorrow at 4pm in Media Lab 054
conceptual
vs Smolensky�s sub-symbolic???