Learning Unbiased Representations via Rényi Minimization

In recent years, significant work has been done to include fairness constraints in the training objective of machine learning algorithms. Differently from classical prediction retreatment algorithms, we focus on fair representations inputs. The challenge is learn that capture most relevant information predict targeted output Y, while not containing any about a sensitive attribute S. We leverage which estimate Hirschfeld-Gebelein-Renyi (HGR) maximal correlation coefficient by deep neural network transformations and use it as min-max game penalize intrinsic bias multi dimensional latent representation. Compared other dependence measures, HGR captures more non-linear dependencies, making algorithm efficient mitigating bias. After providing theoretical analysis consistency estimator its desirable properties for mitigation, empirically study impact at various levels architectures. show acting intermediate architectures provides best expressiveness/generalization abilities using an based loss than adversarial approaches literature.