Seminar on Computational Learning and Adaptation

  The Bias-variance Tradeoff in Active Learning

Hinrich Schuetze
Enkata Technologies

Training set generation is the main cost in many applications of statistical classification.  Active learning is often used for getting the maximum amount of information possible from each labeling decision. However, training sets generated by active learning are biased samples of the underlying population.  How can we compute an unbiased estimate of classifier performance in this scenario? We propose a solution, evaluate it for a text classification problem and discuss the bias-variance tradeoff we face.

Return to the seminar schedule