💡 How much more training is needed to move from a biased to a fair predictor ?
Empirically: up to 400% extra training.

Joint work with François Bachoc, Jérôme Bolte and Jean-Michel Loubes.

(3/3)

- The majority loss dominates the landscape: minimizing the loss is almost always equivalent to minimizing the majority part.

- We introduce a stereotype gap: the distance between fair and majority-optimal critical points.

- Debiasing requires extra training.

(2/3)

📢 New preprint!
“When majority rules, minority loses: bias amplification of gradient descent”

We often blame biased data but training also amplifies biases. Our paper explores how ML algorithms favor stereotypes at the expense of minority groups.

➡️ arxiv.org/abs/2505.13122

(1/3)

📣Job altertS in Toulouse (Maths department) 📣
There are multiple jobs offers from Master internships to Assistant professor in the mathematics of data science, optimization, statistical fairness and robustness.
I will try to regroup them in this thread 🧵