Learning from Failure: Training Debiased Classifier from Biased Classifier

Abstract

Neural networks often learn to make predictions that overly rely on spurious correlation existing in the dataset, which causes the model to be biased. While previous work tackles this issue with domain-specific knowledge or explicit supervision on the spuriously correlated attributes, we instead tackle a more challenging setting where such information is unavailable. To this end, we first observe that neural networks learn to rely on the spurious correlation only when it is "easier" to learn than the desired knowledge, and such reliance is most prominent during the early phase of training. Based on the observations, we propose a failure-based debiasing scheme by training a pair of neural networks simultaneously. Our main idea is twofold; (a) we intentionally train the first network to be biased by repeatedly amplifying its "prejudice", and (b) we debias the training of the second network by focusing on samples that go against the prejudice of the biased network in (a). Extensive experiments demonstrate that our method significantly improves the training of network against various types of biases in both synthetic and real-world datasets. Surprisingly, our framework even occasionally outperforms the debiasing methods requiring explicit supervision of the spuriously correlated attributes.Preprint. Under review.

Federated Learning of Molecular Properties with Graph Neural Networks in a Heterogeneous Setting

“…Learning to reweight training samples are widely used in Curriculum learning (Zhou, Wang, and Bilmes 2020), hard-sample Mining (Lin et al 2017), domain generalization (Sagawa et al 2019;Arjovsky et al 2019;Krueger et al 2021), debiasing (Nam et al 2020), model calibration (Mukhoti et al 2020), adversarial defense (Zhang et al 2020), etc. Our method is closely related to Focal Loss (Lin et al 2017) and worst case optimization (Sagawa et al 2019).…”

Learning Bias-Invariant Representation by Cross-Sample Mutual Information Minimization

“…Recently, several works focus on learning with weak or even no bias supervision [27,35,51,27]. Learning from Failure (LfF) puts more weight on failure samples [35].…”

“…Several methods have been proposed to learn to remove the dataset bias [25,1,41,56,35,5]. Among them, some methods regularize the model to not learn bias with additional regularization terms [35,5], and others learn to eliminate the learned bias information by adversarial learning [25,41,56].…”

“…Several methods have been proposed to learn to remove the dataset bias [25,1,41,56,35,5]. Among them, some methods regularize the model to not learn bias with additional regularization terms [35,5], and others learn to eliminate the learned bias information by adversarial learning [25,41,56]. Our work follows the latter and the bias elimination is often conducted by minimizing the correlation between the extracted features and the bias label.…”

NeuroView: Explainable Deep Network Decision Making

“…Interpretability is essential in many fields to figure out if there are any biases from the DN. For example, in [25], the authors performed experiments on an action recognition dataset where the test accuracy is far from stellar due to the action recognition biases. The Soft VQ is denoted as a sigmoid function applied to the layer.…”