On the Robustness of Active Learning
This work addresses the problem of unreliable active learning applications for machine learning practitioners, highlighting incremental improvements in method robustness.
The paper analyzes the robustness of active learning methods across different classifiers, hyperparameters, and noisy labels, finding biases that lead to suboptimal performance when transferring methods between datasets.
Active Learning is concerned with the question of how to identify the most useful samples for a Machine Learning algorithm to be trained with. When applied correctly, it can be a very powerful tool to counteract the immense data requirements of Artificial Neural Networks. However, we find that it is often applied with not enough care and domain knowledge. As a consequence, unrealistic hopes are raised and transfer of the experimental results from one dataset to another becomes unnecessarily hard. In this work we analyse the robustness of different Active Learning methods with respect to classifier capacity, exchangeability and type, as well as hyperparameters and falsely labelled data. Experiments reveal possible biases towards the architecture used for sample selection, resulting in suboptimal performance for other classifiers. We further propose the new "Sum of Squared Logits" method based on the Simpson diversity index and investigate the effect of using the confusion matrix for balancing in sample selection.