Uniform Stability in Machine Learning Theory

~1 min read

TL;DR. Uniform stability — an algorithm whose output barely changes when you swap one training point — directly implies generalization, without needing VC dimension or Rademacher complexity. This writeup walks through the proofs end to end.

• When I took 9.520 at MIT with the wonderful professors Lorenzo Rosasco and Tomaso Poggio, I made a survey of my own explanation of the concepts and mathematical proofs of Uniform Stability and Generalization in (statistical) Learning Theory.
• Document available here. Enjoy!

---

If you’d like to cite this post:

@misc{miranda2019uniformstability,
  author = {Miranda, Brando},
  title  = {Uniform Stability in Machine Learning Theory},
  year   = {2019},
  month  = {October},
  howpublished = {\url{https://brando90.github.io/brandomiranda/2019/10/31/uniform-stability.html}},
  note   = {Blog post}
}