Active Learning via Reduction

Speaker

John Langford, Yahoo! Research

Abstract

We developed a new class of learning reduction allowing us to convert a black-box supervised learning algorithm into an active learning algorithm capable of introspectively determining which examples to label. This approach has several new useful properties:

(1) It reliably works well across several very different base supervised learning algorithms.
(2) It interpolates to supervised learning.
(3) It allows you to switch the choice of learning algorithm.
(4) It can work with great efficiency, depending on the base learning algorithm.

These benefits address the primary drawbacks of active learning relative to more typical supervised learning algorithms, making active learning a routinely useful standard technology which reduces labeling costs by large factors. This new class of reductions may also have further applications in interactive learning.