“For 2,000 years, we believed logic was the only instrument for solving intellectual problems.
Now, our analysis of machine learning is showing us that to address truly complex problems, we need images, poetry, and metaphors as well.” Vladimir Vapnik.
Learning using privileged information (LUPI) was first introduced by Vapnik and Vashist1
as a framework for learning with additional information at training time that is not available at test time.
LUPI framework seeks to devise learning methods that can utilize this additional information with an emphasis that the
privileged information is not required or more importantly not needed to be inferred at test time.
Thus, this information cannot be simply used as an input to the learner.
1 V. Vapnik and A. Vashist. A new learning paradigm: Learning using privileged information. Neural Networks, pp. 544--557, 2009.
We show that the LUPI framework is applicable to several scenarios that have been studied in computer vision before: attributes, bounding boxes and image tags. By exploiting the insight that privileged information allows us to distinguish between easy and hard examples in the training set, we have introduced RankTransfer, MarginTransfer, and PrivilegedNoise methods. Recently, we are exploring the usage of privileged information for incorporating training label annotations confidence and for performing cross dataset learning.