Abstract

Active learning is a powerful tool when labelling data is expensive, but it introduces a bias because the training data no longer follows the population distribution. We can, in fact, remove this bias using corrective weights based on importance sampling. This has two main consequences: first, we show that the bias is actually useful for active learning, especially with overparameterized models like neural networks; second, this technique enables active testing---a new way of doing model evaluation with limited data.

Notes

• References:

  • On Statistical Bias in Active Learning: How and When To Fix It

  • Active Testing: Sample Efficient Model Evaluation

• Personal website can be found here.