Go to TMLR homepageModel Debiasing by Learnable Data Augmentation
Abstract: Deep Neural Networks are well known for efficiently fitting training data, yet experiencing poor generalization capabilities whenever some kind of bias dominates over the actual task labels, resulting in models learning “shortcuts”. In essence, such models are often prone to learn spurious correlations between data and labels. In this work, we tackle the problem of learning from biased data in the very realistic unsupervised scenario, i.e., when the bias is unknown. This is a much harder task as compared to the supervised case, where auxiliary, bias-related annotations, can be exploited in the learning process. This paper proposes a novel 2-stage learning pipeline featuring a data augmentation strategy able to regularize the training. First, biased/unbiased samples are identified by training over-biased models.
Second, such subdivision (typically noisy) is exploited within a data augmentation framework, properly combining the original samples while learning mixing parameters, which has a regularization effect. Experiments on synthetic and realistic biased datasets show state-of-the-art classification accuracy, outperforming competing methods, ultimately proving robust performance on both biased and unbiased examples. Notably, being our training method totally agnostic to the level of bias, it also positively affects performance for any, even apparently unbiased, dataset, thus improving the model generalization regardless of the level of bias (or its absence) in the data.
Submission Length: Long submission (more than 12 pages of main content)
Previous TMLR Submission Url: https://openreview.net/forum?id=eCG5krscQy
Changes Since Last Submission: In this resubmission, we have addressed the key concerns raised during the previous round of reviews. After having revised the original submission, most of the raised remarks were satisfactorily addressed. In short, terminology problems were clarified (Reviewers jPKv and qXQf), and issues related to the role of Mixup and the “unsupervised” nature of the problem have been fixed. Moreover, we introduced more baselines, ablation analyses, and discussed the validation of the hyper-parameters, $\gamma$ in particular (Reviewer qXQf).
Only Reviewer LxSz still remained with some doubts about our rebuttal, and such criticisms in the end have been summarized by the AE to take his final recommendation.
Below, we report our responses and the corresponding changes made to the manuscript.
$\quad$
### 1. Mixup vs. other augmentation strategies.
This seems to be the main concern stated in the ‘Comment’ section of the Decision by Action Editor (AE) box. The Reviewer LxSz (and AE) questioned the necessity of using mixup and requested clarification on whether other data augmentation methods could be similarly effective in the context of debiasing. We would like to clarify that our method is specifically designed to mix bias-aligned and bias-conflicting samples, guided by estimated bias pseudo-labels, and does so regardless of the semantic task labels.
Standard augmentation techniques (e.g., color jittering, etc.) are not suitable for integration into our framework, as they do not explicitly operate across the bias-aligned/conflicting sample split. Nonetheless, to provide a more complete picture, we have included an ablation study in Table 7 in which we apply standard augmentations across all samples, irrespective of bias pseudo-labels. Table 7 also originally compared the cases of no augmentation, vanilla mixup, and all mixing combinations of samples from the biased/unbiased subsets for all considered levels of bias ratio (from 95% to 99.5%), showing that our proposed approach is the best performing one. A discussion has been added in Section 7.6.
Thanks to this comment we were able to show that our mixing strategy succeeds in obtaining the best performance by learning how to mix samples on the basis of the estimated bias pseudo-labels, also evidencing that standard augmentations or naive mixing modalities are less effective or even detrimental for bias mitigation.
$\quad$
### 2. Comparisons with Reweighting and Resampling Methods.
We would like to clarify that while reweighting and resampling are commonly used for debiasing, they are not augmentation strategies, but rather alternative debiasing approaches. We have a detailed comparison with weighted ERM, where biased and unbiased samples are reweighted based on their estimated pseudo-labels. This comparison provides insights into how our learned mixup-based augmentation compares to other well-known debiasing paradigms.
$\quad$
### 3. Label Mixing
A point raised in the previous review suggested that our method contradicts the claim of not mixing samples with different labels. We believe this concern stems from a misunderstanding of our terminology. In our framework:
- Task (semantic) labels refer to the labels used in the main classification task.
- Bias pseudo-labels refer to our estimation of whether a sample is bias-aligned or bias-conflicting.
Our method performs mixup across different bias pseudo-labels — i.e., between estimated biased and unbiased samples — and does so without considering task labels for samples, which are randomly picked.
This distinction is critical and is now clarified in the manuscript. Table 7 includes an ablation study evaluating all possible combinations of mixup strategies across bias pseudo-label groups, including “same bias”, “different bias”, and “random” sampling.
$\quad$
### 4. Two-Pass Training Requirement
The need for two forward passes — one for bias estimation, the second for debiasing — was noted as a potential drawback. However, this approach is widely adopted in the literature. For example, Just Train Twice (Liu et al., 2021) performs full retraining. Learning from Failure (Nam et al., 2020) and other works (e.g., Teney et al., 2022; Li et al., 2022; Kim et al., 2022a; Lemoine et al., 2018) rely on auxiliary models or two-stage pipelines.
This strategy is a common and necessary consequence of the fact that bias is not observable, and must be estimated. Section 2 of the manuscript clarifies that our approach is in line with standard practices in debiasing research.
$\quad$
More details about the entire review process, discussions and our former revisions can be retrieved in the OpenReview portal.
Assigned Action Editor: ~Vinay_P_Namboodiri1
Submission Number: 5562
Loading