Pinned
How many samples do you need from an unknown distribution in order to train a model with multicalibration error at most epsilon?
Answer: 1/epsilon^3 samples is both necessary and sufficient.
Aaron Roth
3,854 posts
Aaron Roth
@Aaroth
CS prof at Penn. Amazon Scholar at AWS. Author of The Ethical Algorithm (w/ Michael Kearns). I study machine learning, privacy, game theory, and uncertainty.
Philadelphia, PA
Joined May 2007
- Michael Kearns and I wrote a book! Its called "The Ethical Algorithm: The Science of Socially Aware Algorithm Design", and its going to be published by Oxford University Press in the fall. Let me tell you about it! (Thread)
- Now that we are all teaching using Zoom, we can potentially open up our classes to the wider socially distant world. I'm teaching algorithmic game theory: cis.upenn.edu/~aaroth/course… Who wants to learn about the VCG mechanism at 3pm on Tuesday March 24? :-)
- I've been enjoying learning about linear regression. This is a really cool machine learning technique with some really elegant theory --- someone should have taught me about this earlier!
- Here are my notes for Learning in Games and Games in Learning cis.upenn.edu/~aaroth/GamesI… (The class is here and links to youtube recordings for the various lectures: mlgametheory.com). We covered lots of topics, both classic and modern. Here is some of it. 🧵
- Reading through the ML interpretability literature. Question: Why do people think assigning Shapley values to features is a reasonable "explanation"? Yes, they have "rigorous foundations", which is true in the sense that Shapley proved they uniquely satisfy certain axioms: 1/n
- The United States has had a tremendous advantage in science and technology because it has been the consensus gathering point: the best students worldwide want to study and work in the US because that is where the best students are studying and working. 1/
- Machine Learning is really good at making point predictions --- but it sometimes makes mistakes. How should we think about which predictions we should trust? In other words, what is the right way to think about the uncertainty of particular predictions? A thread about new work 🧵
- I can finally tell people honestly that I'm an unqualified professor (but I'm sure many suspected this all along.)
- Every time I go to vote I am newly impressed by our local elementary school's motto.
- We wrote a paper proposing a new relaxation of differential privacy that has lots of nice properties: arxiv.org/abs/1905.02383 It's 85 pages long, so here is the TL;DR. Suppose S is a dataset with your data, and S' is the dataset with your data removed. 1/
- Honored to receive the Hans Sigrist prize!His research aims to incorporate social norms into #algorithms and to better protect privacy: The most highly endowed #prize of the #unibern, the Hans Sigrist Prize, will go to the US computer scientist @Aaroth: sohub.io/qu1m. @Penn @CIS_Penn @Wharton
- Congratulations to Cynthia Dwork for a hugely well deserved Knuth Prize! sigact.org/prizes/knuth/c…
