Fairness Literature

Self-Supervised Learning

Prediction of labels, segments, distortions, colorizations.
Can be used to estimate terrain, depth completion to identify distance to objects while driving.

Want to guarantee stable performance even if the data distribution changes.
Causal graphs to be used with identifying transferable knowledge while keeping fairness constraints (protected attributes).
Datasets: group identifier bias, african american english dialect bias, compas recidivism
False positive bias (“I am a gay man” given high toxicity scores).
Demographic parity (If subjects in protected and unprotected groups have equal probability of being assigned to positive predicted class)
Equality of odds (If subjects in protected and unprotected groups have equal TPR and equal FPR)
Equality of opportunity (If subjects in protected and unprotected groups have equal FNR)
Use an adversarial network which attempts to predict (discriminator with loss based on the above principles)

Gold standard of causal inference is experimentation (randomized, controlled trial)
Combine data from different contexts to figure out causation. Combining datasets can show that the contexts can cause the X factor in the pooled data.
Uses directed graphs

BERT embeddings for pronoun resolution are gender-biased performing more poorly on female pronouns
BERT hate speech classifiers correlate protected groups with hateful language (muslim, black)
De-biased models can be transferred with downstream fine-tuning and remain less biased. Can also transfer this to a different domain.