I’m thrilled to announce that I'll be joining @iitmadras as an Assistant Professor in April 2024!. I’m immensely grateful to my amazing mentors, family, and friends for their unwavering support. (1/4).

I'm hiring MS/PhD students! . Lots of benefits: scholarships & salary top-ups, compute, international research visits! . GATE exemptions for: .* MS/MTech degree holders from India/abroad .* Bachelors in a CFTI (IITs/NITs) . See attached post/reach out!.

RT @ai4bharat: 🚀 AI4Bharat: Advancing Indian Language AI - Open & Scalable! 🇮🇳✨. Over the past 4 years, we at AI4Bharat have been on a miss….

RT @lxeagle17: I'm teaching databases this semester at Berkeley. My students all seem unusually brilliant. Not many go to office hours, and….

Calling all potential MS & PhD students! 📢 . @iitmadras has a solid and growing community of international students. Lots of tuition waivers & scholarships are up for grabs: apply by April 15th!. #Masters #PhD #InternationalStudents #scholarships #IITMadras.

RT @ravi_iitm: Honoured and humbled to be elected an @RealAAAI Fellow this year, along with an illustrious cohort. Congratulations to all!….

We're organising the Deployable AI workshop at AAAI 2025! . We're inviting short (4 page) or long (7 page) papers on AI efficiency, explainability, fairness, privacy, reliability robustness, and social impact!. Deadline: Nov. 22, 2024 AoE. Do submit--it'll be a fun workshop!.

Really looking forward to Yann LeCun's upcoming visit to IITM!.

RT @WSAI_IITM: Very proud of our researchers at @ai4bharat for this achievement!.

RT @rdnowak: Here’s my take on the “mathematical foundations” of machine learning and AI. These course notes cover the basics of statistica….

So why does anti-correlated noise work? . It helps by canceling out past noise that we injected. In low signal directions, the gradients are not able to undo the injected past noise. The proofs reveal that anti-correlations really help there!. (6/6).

This theory is practically useful too! . The proofs yield an analytical expression for noise correlations. Empirically, this outperforms DP-SGD and matches SoTA DP-FTRL (which requires solving a SDP for the correlations)!. (5/n).

Lots of technical goodies!. Typical stochastic gradient proofs get tricky with noise that is correlated across steps. We work in the frequency domain!. The proofs combine ideas from multiple fields: signal processing, stochastic approximation, and DP. (4/n).

Key results:. (1) Strict improvement for simple objectives of mean estimation & linear regression. (2) Linear regression: Anti-correlated noise improves the error from ambient dimension to a problem-dependent effective dimension (= lower bound). Bounds match simulations!. (3/n)

How do we inject noise into SGD for DP training?. Independent noise (DP-SGD) is most common but empirical results show that correlated noise could be better (DP-FTRL). We give the first *provable* gap: (anti-)correlated noise is the way to go for *learning* problems!. (2/n).

Excited to present our work “Correlated Noise Provably Beats Independent Noise for Differentially Private Learning” at #ICLR at poster #217 at 4:30 pm today (Wednesday). Joint w/ @Chris_Choquette @DjDvij, @aroonganesh, @shortstein, Abhradeep Thakurta. (1/n)

More details about each paper in their respective threads. Please come by and say hi (or reach out and I'd be happy to chat)!.

I'll be at #ICLR at these posters:. Correlated Noise Provably Beats Independent Noise for Differentially Private Learning (#217 @ 4:30 pm on Wednesday). Distributionally Robust Optimization with Bias and Variance Reduction (Spotlight, #154 @ 10:45 am on Thursday).

RT @Pavan_KumarGV: In the past 14 yrs of my research group, I have made a few policies for myself, which may be helpful to others:. A 🧵. 1)….

Calling motivated students interested in pursuing MS/PhD in ML/AI, specifically privacy & generative AI!. The research group I'm starting at @iitmadras has openings!. Apply by *Mar 31* directly to @DSAI_IITM or @iitmcse at .

RT @QuantaMagazine: Michel Talagrand has been awarded the Abel Prize, one of the highest honors in mathematics, for applying tools from hig….