Rates of convergence in active learning

TL;DR

This paper investigates how quickly active learning algorithms can reduce generalization error under different noise conditions and introduces an adaptive model selection algorithm that outperforms passive learning in certain scenarios.

Contribution

It provides new theoretical bounds on convergence rates in active learning with noise and proposes an adaptive algorithm for model selection within a hierarchy of classifiers.

Findings

Error rates can be significantly faster than passive learning under certain noise conditions.

The proposed algorithm adaptively converges to the best classifier in a hierarchy.

Sufficient conditions are identified for rapid convergence in active learning.

Abstract

We study the rates of convergence in generalization error achievable by active learning under various types of label noise. Additionally, we study the general problem of model selection for active learning with a nested hierarchy of hypothesis classes and propose an algorithm whose error rate provably converges to the best achievable error among classifiers in the hierarchy at a rate adaptive to both the complexity of the optimal classifier and the noise conditions. In particular, we state sufficient conditions for these rates to be dramatically faster than those achievable by passive learning.