Jul
21
On Transfer Learning
July 21, 2007 |
Definition (from DARPA): The ability of a system to recognize and apply knowledge and skills learned in previous tasks to novel tasks
Current approaches involve either the building of a shared model of a domain or multiple domains, in the form of a case base, hierarchy, or relational schema, that couple the classifiers together, or the creation of mapping between distinct representations. Bayesian and neural approaches dominate the research thus far.
(from Droy 2007-IJCAI07) In spam filtering, a typical data set consists of thousands of labeled emails belonging to a collection of users. In this sense, we have multiple data sets–one for each user. Should we combine the data set and ignore the prior knowledge that different users labeled each email? If we combine the data from a group of users who roughly agree on the definition of spam we will have increased the available training data from which to make predictions. However, if the preferences within a population of users are heterogeneous, then we should expect that simply collapsing the data into an undifferentiated collection will make our predictions worse.
Resources
Caruana dissertation (1997). Part of ALVINN
Berkeley 2005 course. Reading list. Bayesian approaches are focused in on.
Oregon State 2005 course. Probabilistic Relational Models.
DARPA Proposal. Now in its third and final year.
CBR Approach. Strategy game playing.
Wikipedia Entry
Workshops
NIPS 1995
Inductive Transfer : 10 Years Later (2005)
