Learning Adversarially Fair and Transferable Representations

Authors: David Madras, Elliot Creager, Toniann Pitassi, Richard Zemel

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate experimentally that classifiers trained naively (without fairness constraints) from representations learned by our algorithm achieve their respective fairness desiderata; furthermore, we show empirically that these representations achieve fair transfer they admit fair predictors on unseen tasks, even when those predictors are not explicitly specified to be fair.
Researcher Affiliation Academia 1Department of Computer Science, University of Toronto, Toronto, Canada 2Vector Institute, Canada.
Pseudocode Yes ALGORITHM 1 Evaluation scheme for fair classification (Y = Y ) & transfer learning (Y = Y ).
Open Source Code No The paper does not contain any explicit statement about providing open-source code for the methodology or a link to a code repository.
Open Datasets Yes We evaluate the performance of our model on fair classification on the UCI Adult dataset5, We aimed to predict each person s income category (either greater or less than 50K/year).
Dataset Splits Yes We split our test set ( 20, 000 examples) into transfer-train, -validation, and -test sets.
Hardware Specification No The paper does not provide any specific details about the hardware used for running the experiments, such as GPU or CPU models.
Software Dependencies No The paper mentions optimization algorithms like SGD and Adam, but does not provide specific version numbers for any software dependencies or libraries used in the implementation.
Experiment Setup Yes Each sub-figure shows the accuracy-fairness trade-off (for varying values of γ; we set α = 1, β = 0 for all classification experiments); We used L1 loss for the classifier; For transfer learning, we set our reconstruction coefficient β = 1; We can optionally set our classification coefficient α to 0; trained LAFTR (α = 0, β = 1, ℓ2 loss for the decoder) on the full training set.