Revisiting the Importance of Amplifying Bias for Debiasing

In image classification, debiasing aims to train a classifier be less susceptible dataset bias, the strong correlation between peripheral attributes of data samples and target class. For example, even if frog class in mainly consists images with swamp background (i.e., bias aligned samples), debiased should able correctly classify at beach conflicting samples). Recent approaches commonly use two components for debiasing, biased model fB fD. is trained focus on overfitted bias) while fD by concentrating which fails learn, leading bias. While state art techniques have aimed better fD, we training fB, an overlooked component until now. Our empirical analysis reveals that removing from set important improving performance This due fact work as noisy amplifying since those do not include attribute. To this end, propose simple yet effective sample selection method removes construct amplified fB. can directly applied existing reweighting based approaches, obtaining consistent boost achieving both synthetic real-world datasets.