Learning an Ethical Module for Bias Mitigation of pre-trained Models

In spite of the high performance and reliability of deep learning algorithms in broad range everyday applications, many investigations tend to show that a lot of models exhibit biases, discriminating against some subgroups of the population. This urges the practitioner to develop fair systems whose performances are uniform among individuals. In this work, we propose a post-processing method designed to mitigate bias of state-of-the-art models. It consists in learning a shallow neural network, called the Ethical Module, which transforms the deep embeddings of a pre-trained model in order to give more representation power to the disadvantaged subgroups. Its training is supervised by the von Mises-Fisher loss, whose hyperparameters allow to control the space allocated to each subgroup in the latent space. Besides being very simple, the resulting methodology is more stable and faster than most current bias mitigation methods. In order to illustrate our idea in a concrete use case, we focus here on gender bias in facial recognition and conduct extensive numerical experiments on standard datasets.

PDF Abstract
No code implementations yet. Submit your code now

Tasks


Datasets


Results from the Paper


  Submit results from this paper to get state-of-the-art GitHub badges and help the community compare results to other papers.

Methods


No methods listed for this paper. Add relevant methods here