Active Learning with Neural Networks
Holger Schoener · 2007
This paper is a survey over the area of active learning algorithms for supervised Neural Networks. These algorithms are concerned with the selection of informative examples to be included into the training set used for learning. After an introduction into the field and the motives for its development, three practical algorithms will be introduced. These are a network function inversion algorithm, one which searches for separating hyperplanes in input space, and the "Query by Committee" algorithm, which accepts new examples only if they provide new information. The last part of this paper explains a theoretical framework to deal with active learning algorithms, deriving results about their behavior from two examples. 1 Introduction 1.1 Motives and History The phrase "Active Learning" reflects the switch from the passive role of Neural Networks (for an introduction to Neural Networks see e.g. [Hertz91], [Haykin94], or [Bishop95]), which are traditionally presented a database of trainin...