Landauer's vision vision
------------------------
Xiaofeng Mi

What's the term x document matrix for vision?

documents are images
some ideas for terms:
 * individual pixels
 * 3x3 patches (need to discretize, VQ??, how big a patch?)
 * pre-align, then pixels (assuming we can recognize objects)
 * objects
 * pieces of objects?
 * (we don't know the correct set of "primitives")
 * surface features/deep features (function)?
 * filter out unimportant/uninformative features somehow

Serious problem: how do you compare pieces?

Does there need to be a symbolic representation in advance?

Can a system extract building blocks to create larger primitives?

Simon Dennis
------------
David DeVault

Need to learn from experience, not just text.

(see powerpoint slides)

Some questions from my reading:

 * What do lists of highly similar words really tell us?

 * The model learns similarity scores directly, not vector
   representations (right?).  So, what can we actually *do* with such
   a representation?

 * While similarity is an important quantity to be computed by a
   representation, this paper points out two kinds of similarity:
   syntagmatic (run-fast) and paradigmatic (run-trot).  This model
   keeps them separate explicitly.

 * Paper does a nice job of explaining what the similarities mean (by
   comparing to human data).  Doesn't describe the algorithm real
   well, though.