Nowadays an ever-growing concerning phenomenon, the emergence of algorithmic biases that can lead to unfair models, emerges. Several debiasing approaches have been proposed in the realm of deep learning, employing more or less sophisticated approaches to discourage these models from massively employing these biases. However, a question emerges: is this extra complexity really necessary? Is a vanilla-trained model already embodying some ``unbiased sub-networks'' that can be used in isolation and propose a solution without relying on the algorithmic biases? In this work, we show that such a sub-network typically exists, and can be extracted from a vanilla-trained model without requiring additional training. We further validate that such specific architecture is incapable of learning a specific bias, suggesting that there are possible architectural countermeasures to the problem of biases in deep neural networks.

In this paper, we study the "dataset bias" problem from a statistical standpoint, and identify the main cause of the problem as the strong correlation between a class attribute u and a non-class attribute b in the input x, represented by p(u|b) differing significantly from p(u). Since p(u|b) appears as part of the sampling distributions in the standard maximum log-likelihood (MLL) objective, a model trained on a biased dataset via MLL inherently incorporates such correlation into its parameters, leading to poor generalization to unbiased test data. From this observation, we propose to mitigate dataset bias via either weighting the objective of each sample n by \frac{1}{p(u_{n}|b_{n})} or sampling that sample with a weight proportional to \frac{1}{p(u_{n}|b_{n})}. While both methods are statistically equivalent, the former proves more stable and effective in practice. Additionally, we establish a connection between our debiasing approach and causal reasoning, reinforcing our method's theoretical foundation. However, when the bias label is unavailable, computing p(u|b) exactly is difficult. To overcome this challenge, we propose to approximate \frac{1}{p(u|b)} using a biased classifier trained with "bias amplification" losses. Extensive experiments on various biased datasets demonstrate the superiority of our method over existing debiasing techniques in most settings, validating our theoretical analysis.

Neural networks trained with ERM (empirical risk minimization) sometimes learn unintended decision rules, in particular when their training data is biased, i.e., when training labels are strongly correlated with undesirable features. To prevent a network from learning such features, recent methods augment training data such that examples displaying spurious correlations (i.e., bias-aligned examples) become a minority, whereas the other, bias-conflicting examples become prevalent. However, these approaches are sometimes difficult to train and scale to real-world data because they rely on generative models or disentangled representations. We propose an alternative based on mixup, a popular augmentation that creates convex combinations of training examples. Our method, coined SelecMix, applies mixup to contradicting pairs of examples, defined as showing either (i) the same label but dissimilar biased features, or (ii) different labels but similar biased features. Identifying such pairs requires comparing examples with respect to unknown biased features. For this, we utilize an auxiliary contrastive model with the popular heuristic that biased features are learned preferentially during training. Experiments on standard benchmarks demonstrate the effectiveness of the method, in particular when label noise complicates the identification of bias-conflicting examples.

Deep neural networks often make decisions based on the spurious correlations inherent in the dataset, failing to generalize in an unbiased data distribution. Although previous approaches pre-define the type of dataset bias to prevent the network from learning it, recognizing the bias type in the real dataset is often prohibitive. This paper proposes a novel bias-tailored augmentation-based approach, BiaSwap, for learning debiased representation without requiring supervision on the bias type. Assuming that the bias corresponds to the easy-to-learn attributes, we sort the training images based on how much a biased classifier can exploits them as shortcut and divide them into bias-guiding and bias-contrary samples in an unsupervised manner. Afterwards, we integrate the style-transferring module of the image translation model with the class activation maps of such biased classifier, which enables to primarily transfer the bias attributes learned by the classifier. Therefore, given the pair of bias-guiding and bias-contrary, BiaSwap generates the bias-swapped image which contains the bias attributes from the bias-contrary images, while preserving bias-irrelevant ones in the bias-guiding images. Given such augmented images, BiaSwap demonstrates the superiority in debiasing against the existing baselines over both synthetic and real-world datasets. Even without careful supervision on the bias, BiaSwap achieves a remarkable performance on both unbiased and bias-guiding samples, implying the improved generalization capability of the model.