Greg Detre
Tuesday, November 05, 2002
Goldstone�s interested eventually in aligning between directions
he�s quite clear that this is only a toy model, that doesn�t scale
they�re dealing with correspondence, rather than identity or similarity
no good reason to believe that you can model concepts with such small dimensions � y, but it�s a proof of principle
Nick: subtle assumption: ground the similarity � that the distance metric is the same across(/within) systems???
this gets rid of the circularity???
problem with linear scaling??? of dimensions???
inhibition: punish any node that corresponds to >1 node in the other system
most conceptual spaces would have some self-similarity structure � interesting
Hugo couldn�t duplicate the interaction between internal and external
ah, his interaction only works with absolute coordinates
criticisms:
abstracting concepts to conceptual spaces
Einsteinian vs topological spaces
rescale one dimension vs the other, then you can change which points are nearest each other
no, you don�t change relative to each other, but you do change the absolute distances between points
but if you grant that the scaling happens to both system, this isn�t a problem
let�s ignore the issue that it requires mind-reading
problem with identity of distance matrices
does it emerge out of learning algorithms, neural substrates??? but he doesn�t argue for this
ok, what Peter + Nick are saying is that before the external grounding, the internal-only algorithm requires the two systems to be normalised
so, are we happy to allow him a weaker point contra Fodor
there is, I think, a real problem if you don�t have any anchor points to begin with � of course, in the real world, you do have those
there is a lot of work that has to be done, though it can be done, for establishing the distance metric
maybe though you can still undemrine Fodor without implementing this system
the problem is how much mileage they can get against Fodor with their first internal-only algorithm, right???
Deb: would the paper be interesting without the references to Fodor � e.g. CU Boulder aligning vector/spatial semantic relationships across languages (also Jurafsky someone�s Master�s thesis) (latent semantic analysis � common distance metric comes from grounding in frequency, in a defined feature space)
adding the disclaimer sentence (saying exactly what??? re the need/assumption of a common distance metric) seems to show where they�re resolving the circularity
unlabelled dimensions???
Fodor�s claim is just too strong
reduce to a pattern associator???
it doesn�t make any assumptions of orthogonality
does the twin earth example actually make sense???
because what does it mean to make reference to the composition of water being different if your external grounding of it is the same???
presumably, the idea is that two laymen on earth and twin earth would have the same concept of water, even if some scientists on the different planets would have different concepts (because only they knew the difference)
how is this different from two people with marginally different concepts of the same thing???
(e.g. John/Joan: mushrooms grow from spores/seeds)