AmpliBias: Mitigating Dataset Bias through Bias Amplification in Few-shot Learning for Generative Models

Abstract

Deep learning models exhibit a dependency on peripheral attributes of input data, such as shapes and colors, leading the models to become biased towards these certain attributes that result in subsequent degradation of performance. In this paper, we alleviate this problem by presenting AmpliBias, a novel framework that tackles dataset bias by leveraging generative models to amplify bias and facilitate the learning of debiased representations of the classifier. Our method involves three major steps. We initially train a biased classifier, denoted as fb, on a biased dataset and extract the top-K biased-conflict samples. Next, we train a generator solely on a bias-conflict dataset comprised of these top-K samples, aiming to learn the distribution of bias-conflict samples. Finally, we re-train the classifier on the newly constructed debiased dataset, which combines the original and amplified data. This allows the biased classifier to competently learn debiased representation. Extensive experiments validate that our proposed method effectively debiases the biased classifier. © 2023 Copyright held by the owner/author(s). Publication rights licensed to ACM.