Autoencoders can reweight training features to correct proportional bias

Evidence

To deal with such disproportionate data, Amini and colleagues (2019) created an algorithm that combines an original learning task with a variational autoencoder in order to learn the latent structure of the data set. The algorithm then uses the learned latent distributions to reweight certain features during training—a process that can be used to address proportional racial and gender bias (see also Chen, Johansson, & Sontag, 2018). — Joseph Avery, Jim Sidanius, Joel Cooper, Gregory Davis, Amanda N., Bias in the Law, 1187