Proactive Learning is a generalized form of active learning where the learner must reach out to multiple oracles exhibiting different costs and reliabilities (label noise). One of the its major goals is to capture the cost-noise tradeoff in oracle selection. Sequential active learning exhibits coarse accuracy at the beginning and progressively refine prediction at later stages. The ability to learn oracle accuracies over time and select better oracles or oracle ensembles lead to potentially faster error reduction rate as a function of total cost, and thus improve its cost complexity. To realize this potential, we propose a statistical model that adapts to a range of accuracies at different stages of active learning. In a more general scenario, we formulate the problem as maximum submodular coverage subject to a budget envelope. This research is supported by grants from the National Science Foundation.
[1]
Vladimir Vapnik,et al.
Statistical learning theory
,
1998
.
[2]
Maxim Sviridenko,et al.
A note on maximizing a submodular set function subject to a knapsack constraint
,
2004,
Oper. Res. Lett..
[3]
Andrew McCallum,et al.
Piecewise pseudolikelihood for efficient training of conditional random fields
,
2007,
ICML '07.
[4]
Steve Hanneke,et al.
A bound on the label complexity of agnostic active learning
,
2007,
ICML '07.
[5]
Jaime G. Carbonell,et al.
Proactive learning: cost-sensitive active learning with multiple imperfect oracles
,
2008,
CIKM '08.
[6]
John Langford,et al.
Agnostic active learning
,
2006,
J. Comput. Syst. Sci..