Location: Gilman 132
When: October 5th at 1:30 p.m.
Title: Adversarial machine learning and clustered federated learning: a collective dynamics perspective
Abstract: Machine learning and its applications in AI have entered into a new stage in their development: while the use of AI algorithms is widespread, it is imperative to ask how we can guarantee that as these algorithms penetrate into more domains of our lives they will also be sensitive to privacy concerns, make fair decisions, and be both reliable and robust to data corruption. Are we ready to certify when a given algorithm complies with specific requirements and behaves in the way it is intended to?
In this talk, I will discuss 1) adversarial machine learning in supervised learning settings and 2) clustered federated learning, two examples of ML settings where model accuracy is not the sole criterion for training learning systems. I will present novel approaches for the training of models in these two settings that rely on the use of particle dynamics and their analysis. Our solution to the first problem is inspired by the literature of gradient flows in the space of probability measures under the Wasserstein-Fisher-Rao geometry, and our solution to the second problem is inspired by the literature of consensus-based optimization. With this talk I hope to convey the multiple opportunities for mathematicians to participate in the conversation about pressing societal questions in the development of AI models.
Zoom link: https://wse.zoom.us/j/94601022340