1.1 --- /dev/null	Thu Jan 01 00:00:00 1970 +0000
     1.2 +++ b/org/ullman.org	Wed Apr 10 16:38:52 2013 -0400
     1.3 @@ -0,0 +1,145 @@
     1.4 +#+title: Ullman Literature Review
     1.5 +#+author: Robert McIntyre
     1.6 +#+email: rlm@mit.edu
     1.7 +#+description: Review of some of the AI works of Professor Shimon Ullman.
     1.8 +#+keywords: Shimon, Ullman, computer vision, artificial intelligence, literature review
     1.9 +#+SETUPFILE: ../../aurellem/org/setup.org
    1.10 +#+INCLUDE: ../../aurellem/org/level-0.org
    1.11 +#+babel: :mkdirp yes :noweb yes :exports both
    1.12 +
    1.13 +
    1.14 +* Ullman 
    1.15 +
    1.16 +Actual code reuse!
    1.17 +
    1.18 +precision = fraction of retrieved instances that are relevant
    1.19 +  (true-postives/(true-positives+false-positives))
    1.20 +
    1.21 +recall    =  fraction of relevant instances that are retrieved
    1.22 +  (true-positives/total-in-class)
    1.23 +
    1.24 +cross-validation = train the model on two different sets to prevent
    1.25 +overfitting. 
    1.26 +
    1.27 +nifty, relevant, realistic ideas
    1.28 +He doesn't confine himself to unplasaubile assumptions
    1.29 +
    1.30 +** Our Reading
    1.31 +
    1.32 +*** 2002 Visual features of intermediate complexity and their use in classification
    1.33 +
    1.34 +    
    1.35 +
    1.36 +
    1.37 +** Getting around the dumb "fixed training set" methods
    1.38 +
    1.39 +*** 2006 Learning to classify by ongoing feature selection
    1.40 +    
    1.41 +    Brings in the most informative features of a class, based on
    1.42 +    mutual information between that feature and all the examples
    1.43 +    encountered so far. To bound the running time, he uses only a
    1.44 +    fixed number of the most recent examples. He uses a replacement
    1.45 +    strategy to tell whether a new feature is better than one of the
    1.46 +    corrent features.
    1.47 +
    1.48 +*** 2009 Learning model complexity in an online environment
    1.49 +    
    1.50 +    Sort of like the heirichal baysean models of Tennanbaum, this
    1.51 +    system makes the model more and more complicated as it gets more
    1.52 +    and more training data. It does this by using two systems in
    1.53 +    parallell and then whenever the more complex one seems to be
    1.54 +    needed by the data, the less complex one is thrown out, and an
    1.55 +    even more complex model is initialized in its place.
    1.56 +
    1.57 +    He uses a SVM with polynominal kernels of varying complexity. He
    1.58 +    gets good perfoemance on a handwriting classfication using a large
    1.59 +    range of training samples, since his model changes complexity
    1.60 +    depending on the number of training samples. The simpler models do
    1.61 +    better with few training points, and the more complex ones do
    1.62 +    better with many training points.
    1.63 +
    1.64 +    The final model had intermediate complexity between published
    1.65 +    extremes. 
    1.66 +
    1.67 +    The more complex models must be able to be initialized efficiently
    1.68 +    from the less complex models which they replace!
    1.69 +
    1.70 +
    1.71 +** Non Parametric Models
    1.72 +
    1.73 +[[../images/viola-parzen-1.png]]
    1.74 +[[../images/viola-parzen-2.png]]
    1.75 +
    1.76 +*** 2010 The chains model for detecting parts by their context
    1.77 +
    1.78 +    Like the constelation method for rigid objects, but extended to
    1.79 +    non-rigid objects as well.
    1.80 +
    1.81 +    Allows you to build a hand detector from a face detector. This is
    1.82 +    usefull because hands might be only a few pixels, and very
    1.83 +    ambiguous in an image, but if you are expecting them at the end of
    1.84 +    an arm, then they become easier to find.
    1.85 +
    1.86 +    They make chains by using spatial proximity of features. That way,
    1.87 +    a hand can be idntified by chaining back from the head. If there
    1.88 +    is a good chain to the head, then it is more likely that there is
    1.89 +    a hand than if there isn't. Since there is some give in the
    1.90 +    proximity detection, the system can accomodate new poses that it
    1.91 +    has never seen before.
    1.92 +
    1.93 +    Does not use any motion information.
    1.94 +
    1.95 +*** 2005 A Hierarchical Non-Parametric Method for Capturing Non-Rigid Deformations
    1.96 +    
    1.97 +    (relative dynamic programming [RDP])
    1.98 +
    1.99 +    Goal is to match images, as in SIFT, but this time the images can
   1.100 +    be subject to non rigid transformations. They do this by finding
   1.101 +    small patches that look the same, then building up bigger
   1.102 +    patches. They get a tree of patches that describes each image, and
   1.103 +    find the edit distance between each tree. Editing operations
   1.104 +    involve a coherent shift of features, so they can accomodate local
   1.105 +    shifts of patches in any direction. They get some cool results
   1.106 +    over just straight correlation. Basically, they made an image
   1.107 +    comparor that is resistant to multiple independent deformations.
   1.108 +    
   1.109 +    !important small regions are treated the same as nonimportant
   1.110 +     small regions
   1.111 +     
   1.112 +    !no conception of shape
   1.113 +    
   1.114 +    quote:
   1.115 +    The dynamic programming procedure looks for an optimal
   1.116 +    transformation that aligns the patches of both images. This
   1.117 +    transformation is not a global transformation, but a composition
   1.118 +    of many local transformations of sub-patches at various sizes,
   1.119 +    performed one on top of the other.
   1.120 +
   1.121 +*** 2006 Satellite Features for the Classification of Visually Similar Classes
   1.122 +    
   1.123 +    Finds features that can distinguish subclasses of a class, by
   1.124 +    first finding a rigid set of anghor features that are common to
   1.125 +    both subclasses, then finding distinguishing features relative to
   1.126 +    those subfeatures. They keep things rigid because the satellite
   1.127 +    features don't have much information in and of themselves, and are
   1.128 +    only informative relative to other features.
   1.129 +
   1.130 +*** 2005 Learning a novel class from a single example by cross-generalization.
   1.131 +
   1.132 +    Let's you use a vast visual experience to generate a classifier
   1.133 +    for a novel class by generating synthetic examples by replaceing
   1.134 +    features from the single example with features from similiar
   1.135 +    classes.
   1.136 +
   1.137 +    quote: feature F is likely to be useful for class C if a similar
   1.138 +    feature F proved effective for a similar class C in the past.
   1.139 +
   1.140 +    Allows you to trasfer the "gestalt" of a similiar class to a new
   1.141 +    class, by adapting all the features of the learned class that have
   1.142 +    correspondance to the new class.
   1.143 +
   1.144 +*** 2007 Semantic Hierarchies for Recognizing Objects and Parts
   1.145 +
   1.146 +    Better learning of complex objects like faces by learning each
   1.147 +    piece (like nose, mouth, eye, etc) separately, then making sure
   1.148 +    that the features are in plausable positions.