Invariant Representations and the Scanner Problem

Scanner bias is a known source of variation in modern multi-site
imaging studies. Current best practices all use forms of regression,
covarying for site. In this talk I will describe an alternate method
that instead exploits invariant representations and the data
processing inequality, with preliminary results on a multi-site
diffusion MRI dataset.

Along the way I will describe recent results from our group on
learning such invariant representations in a variational setting
(using VAE), implications for adversarial training schema, and other
use cases of invariant representations, such as style transfer and
fair representation.