Assessing and enhancing robustness of active learning strategies to spurious bias
Abstract: We focus on investigating the performance of common active learning (AL) algorithms under spurious bias and designing an AL algorithm that is robust to spurious bias. Spurious bias refers to the bias created when certain potentially simpler, task-irrelevant attributes in the training set are highly correlated with the target labels. Spurious bias can occur if the sample we use for analysis is not representative of the population, some samples are overrepresented while others are underrepresented. The AL criteria share similarities with approaches to
addressing spurious correlations in passive settings. Hence, with an appropriately defined acquisition function, a sample-efficient framework can be established to effectively handle spurious correlations. Inspired by recent works on simplicity bias, we propose Domain-Invariant Active Learning (DIAL) which leverages the disparity in training dynamics between overrepresented and underrepresented samples, selecting samples that exhibit “slow” training dynamics. DIAL involves no excessively resource-intensive computations as it only relies on training checkpoints to estimate the dynamics of the samples, making it more scalable for addressing real-world spurious correlation problems with AL.
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: The revision has been made based on the feedback from the reviewers. All changes are coloured in green in the revised manuscript (except for the minor changes such as typos, formatting, plotting, references, etc).
**Experiments**
1. We found a bug in our implementation of BADGE. The bug has been fixed and the results are
updated. But the results are not significantly different from the previous version.
2. SOTA performance is added (in Table B.2) for the comparison. [Reviewer bAF9 ]
3. Extra baselines are added in the experiments. [Reviewer zGQQ ]
**Clarification, explanation, and discussion**
1. The statements about constant learning rate and learning rate scheduler are clarified. [Reviewer zGQQ, kRh4, oT5d]
2. Abstract and introduction are shortened and improved [Reviewer bAF9, oT5d].
3. An example of Coloured MNIST is added in the introduction for better problem understanding. [Reviewer kRh4]
4. The concern of balanced query batch and synthetic example are clarified in Section 4 and Section 5.3 [Reviewer zGQQ ].
5. The information of the number of checkpoints and the distance function used in the experiments are added in Section 7 [Reviewer kRh4].
6. Limitations and future work are added in the conclusion [Reviewer kRh4, oT5d].
7. Broader impact statement is added [Reviewer kRh4].
**Minor issues**
1. bibliography, typos, formatting, terminology, and plotting issues are fixed [All reviewers].
Assigned Action Editor: ~Jaakko_Peltonen1
Submission Number: 2943
Loading