Learning Unbiased Representations via Mutual Information Backpropagation

Abstract

We are interested in learning data-driven representations that can generalize well, even when trained on inherently biased data. In particular, we face the case where some attributes (bias) of the data, if learned by the model, can severely compromise its generalization properties. We tackle this problem through the lens of information theory, leveraging recent findings for a differentiable estimation of mutual information. We propose a novel end-to-end optimization strategy, which simultaneously estimates and minimizes the mutual information between the learned representation and the data attributes. When applied on standard benchmarks, our model shows comparable or superior classification performance with respect to state-of-the-art approaches. Moreover, our method is general enough to be applicable to the problem of "algorithmic fairness", with competitive results.

Anatomizing Bias in Facial Analysis

“…(Tartaglione, Barbano, and Grangetto 2021) proposed a regularization strategy to disentangle the biased features while entangling the features belonging to the same target class. Apart from this some techniques are proposed that use feature distillation (Jung et al 2021) and mutual information between the learned representation (Ragonesi et al 2021). For understanding different bias mitigation algorithms (Wang et al 2020) provided a thorough analysis of the existing bias mitigation techniques.…”

Information-Theoretic Bias Reduction via Causal View of Spurious Correlation

“…To address this issue, numerous debiasing frameworks have been proposed for identifying and mitigating the potential risks posed by dataset or algorithmic biases. These frameworks can be categorized into pre-processing (Li and Vasconcelos 2019;Sagawa et al 2020b;Kamiran and Calders 2012), in-processing (Sagawa et al 2020a;Sohoni et al 2020;Wang et al 2019;Zhang, Lemoine, and Mitchell 2018;Gong, Liu, and Jain 2020;Seo, Lee, and Han 2021;Ragonesi et al 2021;Wang et al 2020;Guo et al 2020), and post-processing (Hardt, Price, and Srebro 2016;Zhao et al 2017) (Chu, Kim, and Han 2021;Gong, Liu, and Jain 2020;Ragonesi et al 2021), or robust optimization (Sagawa et al 2020a;Sohoni et al 2020;Seo, Lee, and Han 2021). Postprocessing methods modify the predicted outputs to meet fairness criterion, mainly by calibrating the outputs (Hardt, Price, and Srebro 2016;Zhao et al 2017).…”

“…Our debiasing framework belongs to in-processing methods since it adopts a bias regularization loss during training for reducing the algorithmic bias. The bias regularization loss is similar to (Ragonesi et al 2021), but the major difference is that we introduce conditional mutual information inspired by the structural causal model. Note that the use of unconditional mutual information as a bias regularizer has a critical limitation in learning the relationship between features and target variables for obtaining debiased models.…”

Learning Debiased Models with Dynamic Gradient Alignment and Bias-conflicting Sample Mining

“…The classic approach is reweighting/resampling training samples based on sample number or loss of different groups [25,26,35,36], or even synthesizing samples from minority groups [1]. Another large group of strategies aims at disentangling bias and intrinsic cues in feature domain [17,29,34,41,53], e.g., EnD [41] designs regularizers to disentangle representations with the same bias label and entangle features with the same target label; and some other studies learn disentangled representation by mutual information minimization [17,34,53]. Besides, Sagawa et al [35] and Goel et al [8] aim to improve the worst-group performance by group distributionally robust optimization [9] and Cycle-GAN [52] based data augmentation, respectively.…”