A${}^2$: Adaptive Augmentation for Effectively Mitigating Dataset Bias

Recently, deep neural networks (DNNs) have become the de facto standard to achieve outstanding performances and demonstrate significant impact on various computer vision tasks for real-world scenarios. However, the trained networks can often suffer from overfitting issues due to the unintended bias in a dataset causing inaccurate, unreliable, and untrustworthy results. Thus, recent studies have attempted to remove bias by augmenting the bias-conflict samples to address this challenge. Yet, it still remains a challenge since generating bias-conflict samples without human supervision is generally difficult. To tackle this problem, we propose a novel augmentation framework, Adaptive Augmentation (A${}^2$), based on a generative model that help classifiers learn debiased representations. Our framework consists of three steps: 1) extracting bias-conflict samples from a biased dataset in an unsupervised manner, 2) training a generative model with the biased dataset and adapting the learned biased distribution to the extracted bias-conflict samples’ distribution, and 3) augmenting bias-conflict samples by translating bias-align samples. Therefore, our classifier can effectively learn the debiased representation without human supervision. Our extensive experimental results demonstrate that A${}^2$ effectively augments bias-conflict samples, mitigating widespread bias issues. The code is available in here (https://github.com/anjaeju/A2-Adaptive-Augmentation-for-Effectively-Mitigating-Dataset-Bias).