Flexibly Fair Representation Learning by Disentanglement
summary
Proposes a single encoder that outputs a Z and B latent split (Z for non‑sensitive info, B for per‑attribute latents), trained with VAE reconstruction plus supervised terms that enforce predictiveness of specific B dimensions and independence via an alpha/gamma penalty. At test time the system reconstructs Z and masks selected B coordinates (masking B to remove attributes) before training downstream classifiers. Evaluated on synthetic shapes, CelebA and community‑crime data: works well when attributes are independent, shows accuracy/fairness tradeoffs and degrades under strong correlations; sometimes beats baselines but not universally.
AI-written summary · about
chapters (8)
- 0:00Start
- 0:51Why Fair Representation Learning?
- 3:08Motivation: Subgroup discrimination
- 6:16Motivation: Disentangled Representation Learning
- 23:20Experimental Setup
- 31:31Results - Synthetic Data Audits
- 32:21Results - Tabular and Image Data
- 34:47Conclusion
related talks
42mBeyond Adversarial Approaches on Distributional Robustness
13mCan AI design a fairer future?
30mTalk by Suhas Diggavi (UCLA)
01h 01mDistribution Shift as Underspecification, and What We Might Do About It
28mPosterior Collapse in Deep Generative models
34mInterpretability and Algorithmic Fairness
57mTutorial on Deep Learning III
16mBLens: Contrastive Captioning of Binary Functions using Ensemble Embedding