Chris Russell
@cruss.bsky.social
Algorithmic governance and computer vision. Assoc Prof Oxford, Ellis fellow
The trick is model surgery on a validation set. We train a multi-head model, the first head solves the original task and the other heads predict groups using a squared loss. A weighted sum of all these heads can enforce any fairness definition, and has the same architecture as the original net
December 11, 2024 at 11:29 AM
The trick is model surgery on a validation set. We train a multi-head model, the first head solves the original task and the other heads predict groups using a squared loss. A weighted sum of all these heads can enforce any fairness definition, and has the same architecture as the original net
New fairness toolkit at #NeurIPS today. This fixes most of the problems I've run into in the field.
It is robust to overfitting, works for #NLP and computer vision, and can enforce any definition of fairness that can be written as a function of a confusion matrix. t.ly/ZpRJ-
How do we do that....
It is robust to overfitting, works for #NLP and computer vision, and can enforce any definition of fairness that can be written as a function of a confusion matrix. t.ly/ZpRJ-
How do we do that....
December 11, 2024 at 11:29 AM
New fairness toolkit at #NeurIPS today. This fixes most of the problems I've run into in the field.
It is robust to overfitting, works for #NLP and computer vision, and can enforce any definition of fairness that can be written as a function of a confusion matrix. t.ly/ZpRJ-
How do we do that....
It is robust to overfitting, works for #NLP and computer vision, and can enforce any definition of fairness that can be written as a function of a confusion matrix. t.ly/ZpRJ-
How do we do that....