one shot learning via compositions of meaningful patches ccvlweb.cs.ucla.edu › ~alexw ›...
TRANSCRIPT
One Shot Learning via Compositions of Meaningful Patches CCVL
Alex Wong Alan L. Yuille University of California, Los Angeles
http://ccvl.stat.ucla.edu/ !
Current state-‐of-‐the-‐art algorithms perform very well on most common datasets when trained on thousands of examples However, humans are able to learn a concept from very few
examples, perhaps even just one
mo8va8on
One shot learning is an object categoriza@on task where very few examples (1-‐5) are given for training
what is one shot learning?
• Learn a meaningful patch-‐based representa@on of the underlying structure of an object without human supervision
• Build a composi@onal model composed of a set of compact dic@onaries of meaningful patches
• Reconstruct the target image with deforma@ons of the meaningful patch dic@onaries by patch matching
• Select the class of the best proposed reconstruc@on as label
our approach
Training Images!Segmented Images into Parts!
(Features)!
Compositional Model!
Test Image!Reconstruction!
Select Best Fit!
Parts Representing !Each Region!
Dictionaries of Deformed Parts!
• Our composi@onal model outperforms popular algorithms on the recogni@on task under one shot learning
• The extracted features are seman@cally meaningful • The model generalizes beyond the training set and demonstrates
transferability between separate datasets
conclusion
experimental results
• Symmetry axis acts as a robust object descriptor
• Branch points separate one meaningful part from another • Small segments are merged with nearby meaningful parts
feature extrac8on
• Similar parts, defined by a high match score via Normalized Cross Correla@on, are merged to create a compact dic@onary
• An AND-‐OR graph of the part rela@ons is construc@on for m patches for samples t and u:
• Deforma@ons are applied to the meaningful patches
composi8onal model
We would like to thank Brian Taylor for performing experiments and edi@ng the manuscript. This work was supported by NSF STC award CCF-‐1231216 and ONR N00014-‐12-‐1-‐0883.
acknowledgements
sample reconstruc8ons
truth=1! label=2!
label=6!truth=1!
truth=2! label=7!
truth=2! label=4!
truth=3! label=2!
truth=3! label=5! truth=4! label=9!
truth=4! label=9! truth=5! label=4!
truth=5! label=3!
truth=6! label=4!
truth=6! label=5!
truth=7! label=3!
truth=7! label=2! truth=8! label=6!
truth=8! label=2!
truth=9! label=8!
truth=9! label=4! truth=0! label=3!
truth=0! label=9!
MNIST USPS (trained on MNIST) Images in the Wild (trained on MNIST) Misclassifications
*left image denotes test image, right image denotes reconstruction