Model Debiasing by Learnable Data Augmentation

TL;DR

This work tackles unsupervised debiasing by first identifying biased versus unbiased samples without bias annotations, then learning unbiased representations through a learnable data augmentation that mixes biased and unbiased examples. The bias identification stage offers two schemes, SP and PH, with prediction-history (PH) delivering more reliable separation. The second stage introduces a learnable mixup (l-mix) that optimizes mixing parameters via a neural predictor and an adversarial gradient reversal, yielding a strong regularization effect. Across synthetic and real-bias benchmarks, the method achieves state-of-the-art or competitive accuracy on unbiased test sets while maintaining performance on biased data, demonstrating robust generalization even when bias levels vary or are absent.

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.

Figures (9)

• Figure 1: Problem description. (a) Biased datasets exhibit a class/domain imbalance, namely, one or more classes are mostly observed under one domain, leaving other options under-represented. In the case of supervised debiasing, additional information/annotations regarding the domain distribution are available. (b) In unsupervised debiasing settings, only class labels are available. Possibly, pseudo-labelling can be adopted to fall back into the supervised settings scenario. (c) Biased samples are fitted more easily than unbiased ones, as indicated by the different rates at which the average loss decreases. (Best viewed in color).
• Figure 2: Training accuracy and Pearson correlation coefficient $\rho$ evolution over time. $\rho$ is computed between ground truth labels for $\mathcal{D}_{bias}$, $\mathcal{D}_{unbias}$ and the binary correct/incorrect vector $\mathbf{s}$. Similarly we compute $\rho$ for bias/unbias and the loss values per sample.
• Figure 3: Binary vector $\mathbf{S}$ on the left: rows represent samples indexes $i$ while columns represent epochs $t$. $s_i^t$ is $0$ when sample $i$ is misclassified at epoch $t$ or $1$ when it is correctly classified. When summing up along the epochs, we get the vector $\hat{s} = \sum_{t=1}^{K} s^t$ (middle) that represent the "prediction's history" of each sample after $K$ epochs. When computing its histogram (right), we observe that the distribution is severely skewed towards easy samples (those correctly predicted most of the epochs). We highlighted in red the leftmost bin which is significantly higher than its neighbors: this shows how samples that are never correctly classified are more than those that are rarely correctly classified.
• Figure 4: Gradients for $\mathcal{L} (\hat{\mathcal{D}}_{bias}, f_{\theta})$ and $\mathcal{L} (\hat{\mathcal{D}}_{unbias}, f_{\theta})$ are evaluated to produce the weighted ERM contribution. The regularization step using mixed data aims at producing an additional gradient contribution that decreases the loss function for $\hat{\mathcal{D}}_{mix}$ as for Equation \ref{['eq:loss_method1']}. (Best viewed in color)
• Figure 5: Performance of a model trained with augmented samples generated by Mixup for different pairs of parameters $(\alpha, \beta)$ (equations \ref{['eq:mixup']} and \ref{['eq:loss_method1']}). Each dot represents a single run done with parameters of corresponding coordinates. The right color bar represents the model's test accuracy. We show 3 examples for different probability density functions corresponding to specific pairs $(\alpha, \beta)$.