Generating taxonomies

Target

  CVPR 2007 ?

Final Paper Should Cover These Topics

1. How to make taxonomic trees?
   1. Phyologenetic literature
   2. Just define inter-class distances using the confusion matrix
   3. Various algorithms and their speed/accuracy tradeoffs
2. Why taxonomic trees?
   1. Intrinsically interesting
   2. Related to clustering and spectral embedding [1]
   3. Useful for nested top-down category tests
      1. This is it's own topic
3. Potentially uses
   1. Benchmarking
      1. Guessing that a dog is breadmaker is not as bad as guessing a "dog" is a goat
      2. This suggests a more useful metric than the confusion matrix (Alex)
   2. Guessing about hard-to-classify objects. Discriminate between:
      1. Unknown interesting objects (worth learning or asking about)
      2. Unknown uninteresting objects (clutter, not worth learning)
   3. The above discrimation an important part of how a computer (or child) learns
      1. Relationships between categories
      2. Can ask informative questions about its environment
      3. Closes the feedback loop
      4. Unlimited learning?
         1. But we'll hit restrictions
         2. Exhaustive labelling impossible
         3. Memory limited
            1. see Ryan

Experiment #1

1. Learning
   1. Use n1 categories from Caltech 101 / 256
2. Testing
   1. n2 images from those same categories
   2. n3 images from different categories
   3. clutter
3. What do we expect?
   1. Clutter activates very remote leaves on the tree
      1. How exactly do we define "uninteresting"?
   2. Interesting objects activate nearby leaves
      1. What did we guess?
4. How do we ask a minimal set of questions about training data
   1. Sort out remainder of tree with minimal little supervision
      1. Ask Ryan/Alex/Pietro for ideas?