A VAE trained on dermatology images reconstructs lighter skin better than darker skin even under balanced training, and its uncertainty estimates fail to warn about this disparity.
Science advances 8(31), eabq6147 (2022)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Are generative models fair? A study of racial bias in dermatological image generation
A VAE trained on dermatology images reconstructs lighter skin better than darker skin even under balanced training, and its uncertainty estimates fail to warn about this disparity.