Assistant Professor of ECE and ISyE at Georgia Tech. Interested in anything involving game theory, statistical and/or online learning. Co-organizer of @let4all.
Atlanta, GA, USA
Joined January 2024
- Tweets77
- Following99
- Followers147
- Likes81
Pinned Tweet
I was going to post about my pictured now-graduated students, but Tyler beat me to it with his ever-present enthusiasm :)
This was a really special start to the summer for me, having the privilege of having graduated 3 fantastic PhD students. Post about each of them to follow.
Overdue update: I've defended my dissertation and graduated from @GeorgiaTech! Next, I've joined @MSFTResearch @ms_aifrontiers as a Senior Researcher working on future capabilities of multimodal agents 🚀
Vidya Muthukumar retweeted
In September, we hosted a working group to think about how the TCS community should respond to AI progress. Their report, released yesterday, highlights 12 near-term actions that received broad support among the participants.
simons.berkeley.edu/news-pub…
Vidya Muthukumar retweeted
Why doesn’t OpenAI use an equal amount of the money they spend to prove interesting conjectures in funding PhDs in the mathematical sciences?
That way we will have a next generation of researchers that can understand these results and formulate the next wave of conjectures.
Vidya Muthukumar retweeted
Very honored to be named a Damon Runyon postdoctoral fellow in the Bertozzi lab! Grateful for the support of all my mentors and colleagues throughout this process.
damonrunyon.org/discovery/da…
(1/3) Re-upping this thread because I ended up not making it to COLT 😑. @kuoweilai did a great job leading this work and will be presenting it today afternoon in sunny San Diego!
Change of flavor again -
arxiv.org/abs/2603.04895 (COLT'26)
How does GD favor certain global minimum out of many, when training a shallow ReLU neural network to regress high-dim data (under certain conditions, of course)?
(2/3) in a nutshell, we show that: a) GD dynamics become surprisingly tractable with high-dimensional (but not exactly orthogonal) data on shallow ReLU regression, b) the implicit bias becomes additively close to, but not equivalent to, the min-norm solution.
(1/3) Check out Tyler's thread below! He did a wonderful job leading this year-long, challenging project, and we're very excited about it. A few follow-up comments/shout-outs to work that we built on below.
New paper — the last and best from my PhD!
We show the first end-to-end feature learning analysis of how neural networks overfit to spurious correlations. Our theory implies phase transitions in SGD learning dynamics that match empirical results.
Paper: arxiv.org/abs/2606.30444
(3/3) Our theory partially corroborates interesting findings by @SurbhiGoel_ on synthetic Boolean datasets regarding emergence and relative presence of spurious features and core features. There remain many open questions here that we're excited to study.
(2/3) Our data model and analysis was directly inspired by @MargalitGlasgow's great work on showing that online minibatch SGD learns the XOR model fast. (But the spurious correlation makes analysis *very* different and challenging over and above her work!)
I have really enjoyed referring to this book when teaching my online learning class!
Final version of my book (with a new title)
Online Learning: A Modern Introduction Using
Convex Optimization
Especially proud of the Foreword by @NicoloCB!
It'll be printed by Cambridge University Press.
The end of 7 years of updates :)
arxiv.org/pdf/1912.13213
Vidya Muthukumar retweeted
We are excited to share a new interview as part of the podcast series "Probably Approximately Correct Learners" conducted by
@charapod
- featuring OR expert @hamsabastani
! You can access the podcast *for free* on YouTube: youtube.com/watch?v=lRlcz9cD… any of your favorite platforms.
And thanks to awesome co-advisors @Thejakeyboy and Justin Romberg, it has been a real pleasure. :)
(3/3) Last but not least, presenting: Tyler LaBonte @tmlabonte, who broadened my horizons with great empirical and theoretical work across diverse aspects of OOD generalization!
(2/3) Presenting: Kuo-Wei Lai, who conducted pioneering work on the implicit bias of general loss functions in linear models, the implicit bias of shallow ReLU networks on high-dimensional data (to be presented at COLT next week) and OOD generalization analysis.
Learn more about Kuo-Wei's work at his website: kuoweilai.com/ also, @kuoweilai will be at COLT with me next week!