#+title: Ullman Literature Review

 #+author: Robert McIntyre

 #+email: rlm@mit.edu

 #+description: Review of some of the AI works of Professor Shimon Ullman.

 #+keywords: Shimon, Ullman, computer vision, artificial intelligence, literature review

 #+SETUPFILE: ../../aurellem/org/setup.org

 #+INCLUDE: ../../aurellem/org/level-0.org

 #+babel: :mkdirp yes :noweb yes :exports both

 

 

 * Ullman 

 

 Actual code reuse!

 

 precision = fraction of retrieved instances that are relevant

   (true-positives/(true-positives+false-positives))

 

 recall    =  fraction of relevant instances that are retrieved

   (true-positives/total-in-class)

 

 cross-validation = train the model on two different sets to prevent

 overfitting, and confirm that you have enough training samples.

 

 nifty, relevant, realistic ideas

 He doesn't confine himself to implausible assumptions

 

 ** Our Reading

 

 *** 2002 Visual features of intermediate complexity and their use in classification

 

     

 

 

     Viola's PhD thesis has a good introduction to entropy and mutual

     information 

 

 ** Getting around the dumb "fixed training set" methods

 

 *** 2006 Learning to classify by ongoing feature selection

     

     Brings in the most informative features of a class, based on

     mutual information between that feature and all the examples

     encountered so far. To bound the running time, he uses only a

     fixed number of the most recent examples. He uses a replacement

     strategy to tell whether a new feature is better than one of the

     current features.

 

 *** 2009 Learning model complexity in an online environment

     

     Sort of like the hierarchical Bayesan models of Tennanbaum, this

     system makes the model more and more complicated as it gets more

     and more training data. It does this by using two systems in

     parallel and then whenever the more complex one seems to be

     needed by the data, the less complex one is thrown out, and an

     even more complex model is initialized in its place.

 

     He uses a SVM with polynomial kernels of varying complexity. He

     gets good performance on a handwriting classification using a large

     range of training samples, since his model changes complexity

     depending on the number of training samples. The simpler models do

     better with few training points, and the more complex ones do

     better with many training points.

 

     The final model had intermediate complexity between published

     extremes. 

 

     The more complex models must be able to be initialized efficiently

     from the less complex models which they replace!

 

 

 ** Non Parametric Models

 

 [[../images/viola-parzen-1.png]]

 [[../images/viola-parzen-2.png]]

 

 *** 2010 The chains model for detecting parts by their context

 

     Like the constellation method for rigid objects, but extended to

     non-rigid objects as well.

 

     Allows you to build a hand detector from a face detector. This is

     useful because hands might be only a few pixels, and very

     ambiguous in an image, but if you are expecting them at the end of

     an arm, then they become easier to find.

 

     They make chains by using spatial proximity of features. That way,

     a hand can be identified by chaining back from the head. If there

     is a good chain to the head, then it is more likely that there is

     a hand than if there isn't. Since there is some give in the

     proximity detection, the system can accommodate new poses that it

     has never seen before.

 

     Does not use any motion information.

 

 *** 2005 A Hierarchical Non-Parametric Method for Capturing Non-Rigid Deformations

     

     (relative dynamic programming [RDP])

 

     Goal is to match images, as in SIFT, but this time the images can

     be subject to non rigid transformations. They do this by finding

     small patches that look the same, then building up bigger

     patches. They get a tree of patches that describes each image, and

     find the edit distance between each tree. Editing operations

     involve a coherent shift of features, so they can accommodate local

     shifts of patches in any direction. They get some cool results

     over just straight correlation. Basically, they made an image

     comparator that is resistant to multiple independent deformations.

     

     !important small regions are treated the same as unimportant

      small regions

      

     !no conception of shape

     

     quote:

     The dynamic programming procedure looks for an optimal

     transformation that aligns the patches of both images. This

     transformation is not a global transformation, but a composition

     of many local transformations of sub-patches at various sizes,

     performed one on top of the other.

 

 *** 2006 Satellite Features for the Classification of Visually Similar Classes

     

     Finds features that can distinguish subclasses of a class, by

     first finding a rigid set of anchor features that are common to

     both subclasses, then finding distinguishing features relative to

     those subfeatures. They keep things rigid because the satellite

     features don't have much information in and of themselves, and are

     only informative relative to other features.

 

 *** 2005 Learning a novel class from a single example by cross-generalization.

 

     Let's you use a vast visual experience to generate a classifier

     for a novel class by generating synthetic examples by replacing

     features from the single example with features from similar

     classes.

 

     quote: feature F is likely to be useful for class C if a similar

     feature F proved effective for a similar class C in the past.

 

     Allows you to transfer the "gestalt" of a similar class to a new

     class, by adapting all the features of the learned class that have

     correspondence to the new class.

 

 *** 2007 Semantic Hierarchies for Recognizing Objects and Parts

 

     Better learning of complex objects like faces by learning each

     piece (like nose, mouth, eye, etc) separately, then making sure

     that the features are in plausible positions.