Controlling Directions Orthogonal to a ClassifierDownload PDF


Sep 29, 2021 (edited Nov 22, 2021)ICLR 2022 Conference Blind SubmissionReaders: Everyone
  • Keywords: orthogonal classifier, invariance
  • Abstract: We propose to identify directions invariant to a given classifier so that these directions can be controlled in tasks such as style transfer. While orthogonal decomposition is directly identifiable when the given classifier is linear, we formally define a notion of orthogonality, and when it exists, in the non-linear case. We also provide a surprisingly simple method for constructing the orthogonal classifier (a classifier utilizing directions other than the given classifier). Empirically, we present three use cases where controlling orthogonal variation is important: style transfer, domain adaptation, and fairness. The orthogonal classifier enables partial style transfer, improves domain adaptation with label shifts, and mitigates the unfairness as a predictor.
  • One-sentence Summary: We develop a notion of orthogonality in classifier, and the corresponding construction and utility.
  • Supplementary Material: zip
13 Replies