Biased, noisy labels & how to fix them (sort of)

less than 1 minute read

Published:

This post is adapted from “From Biased Selective Labels to Pseudo-Labels: An Expectation-Maximization Framework for Learning from Biased Decisions” (ICML 2024).

In the previous post, we discussed the assumption that patients who did not receive a test are negative, and what can go wrong if we make this assumption. So, now what?

Imagine a magical algorithm…

Expectation-maximization, explained

Can we use EM to “repair” estimates for biased labels?