When ~24, C.R Rao did this even before he got his PhD :) He already published Cramér–Rao bound, Rao metric and Rao–Blackwellisation that are widely studied across the books and used even today. A motivation for generations: Phd (1948). some results (1945)!
When 25, C.R Rao already published Cramér–Rao bound, Rao metric and Rao–Blackwellisation, even before he got his PhD :) . motivation & food for thought today eh..when publishing rates are on steroids often with misguided incentive of 'where' they publish as opposed to 'what'.
@shortstein
ah was waiting for
#NeurIPS2025
where the reviews are first submitted independently, followed by the papers. The Toronto matching system would match the reviews with the papers, for instantaneous decisions.
I’m on the faculty job market with a focus on distributed ML/statistical inference and privacy. If you have relevant positions, let me know. Please share/RT.
Won another best student paper award at FL-IJCAI'22. See you in Vienna. 'Visual Transformer Meets CutMix for Improved Acc, Comm Eff, and Data Priv in Split Learning'
Congratulations to Camera Culture researchers
@proneat
, Abhishek Singh, Emily Zhang, Otkrist Gupta, and
@raskarmit
on winning the Mukh Best Paper Runner Up Award at
@IEEEFG2021
!
At the most recent count, MBZUAI faculty, researchers and students have 46 papers accepted at the 12th International Conference on Learning Representations (ICLR) 2024; that’s more than double the number of MBZUAI papers accepted in 2023.
this is so beautiful, they bound reverse KL based on fwd KL<eps for small eps's and multivariate Gaussians. I think
@MountainOfMoon
or
@ccanonne_
would find this wonderful too..
Our DP work accepted to
#AAAI2022
allows for sharing supervised manifold embeddings of deep features from any network with differential privacy with
@julballa
,
@raskarmit
Our work on private Fréchet Means got into TMLR with
@SaitejaUtpala
and
@ninamiolane
@geomstats
"Differentially Private on the Manifold of Symmetric Positive Definite (SPD) Matrices with log-Euclidean Metric"
Happy, our paper "Effects of Privacy-Inducing Noise on Welfare and Influence of Referendum Systems" -accepted to Foundations of Responsible Computing-FORC 2024. We connect Randomized Response for diff privacy to influence & welfare in social choice/voting!
Happy to share our work "Privacy-Welfare Trade-off and Privacy-Influence Scaling Law" . Bridges fields: Differential Privacy & Social Choice Theory. We often hear of privacy-utility tradeoffs. We formalize newer trade-offs & relation btwn Voting & Privacy!
In 1945, published his first paper on what is now known as the Fisher–Rao metric, which introduced differential geometry into statistical inference. This led to opening up the burgeoning field now called information geometry.
Accepted to
#NeurIPS2023
! Posthoc privacy guarantees for collaborative inference with modified Propose-Test-Release with
@tremblerz
,
@raskarmit
. See you at New Orleans.
uniform sampling methods from convex bodies are a boon! hit-and-run-walk, ball walk, billiard walk and so on: here's a simple example run for a ball walk before the machinery is put to use in a real project.
@krikamol
:) not even one reviewer responded on our rebuttal with theoretical details and extended expts-tmrw being last day. This tells me how they wrote their reviews in the first place. I guess 15 mins before their review deadline.
Formal Privacy Guarantees for Neural Network queries by estimating local Lipschitz constant . See you at the ICML Workshop on Formal Verification of Machine Learning!
@tremblerz
@raskarmit
The "Privacy-Welfare Trade-Off": We bridge Differential Privacy with Influence & Welfare of Social Choice using Fourier Analysis of Boolean Fns & first principles. New trade-off after priv-util & priv-fairness tradeoffs! with Ibrahim Suat &
@raskarmit
.
1945 Published first paper from Kolkata (ISI), which led to Cramér–Rao bound, Rao metric & Rao–Blackwellisation
1947 Published foundational work on orthogonal array
1948 Published foundational papers on multivariate anove, multivariate significance tests.
1948 Awarded PhD
The beauty of hodge decomposition, is that it can either be seen from just the lens of vanilla linear algebra, or with interpretations via differential forms, differential operators, groups, simplicial complexes and more!
"AirMixML: Over-the-Air Data Mixup for Inherently Privacy-Preserving Edge ML" has got accepted to IEEE GLOBECOM 2021. 1'st step & POC for differential privacy preserving mixup of data for edge ML via perturbs induced in channel higher util mechs in works..
After completing his PhD in 1948 under the supervision of Ronald Fisher, he published papers on orthogonal arrays, multivariate anova, tests of significance in multivariate analysis in late 40's. Authored a breakthrough book Linear Statistical Inference & its Applic's in 1965.
Our SplitLoRA work for efficient LLM fine-tuning is out! SplitLoRA: A Split Parameter-Efficient Fine-Tuning Framework for Large Language Models. Project Page, Results and Paper at:
I’m co-organizing the 2023 MBZUAI Workshop on Collaborative Learning on Dec 9-11, 2023. It is shaping up to be an exciting cross-disciplinary event! 2nd MBZUAI Workshop on Collaborative Learning. Remote registration will open soon! Mark your calendars.
Probability of cat becoming a dog in terms of ratios of areas of spherical caps. i.e if inner prod <x,y> is > 0 what is probability of <R^Tx, R^Ty> to be < 0 after random projection R? Thats what this screenshot gives
@sp_monte_carlo
Finding such kernel embeddings is equivalent to solving a least squares under a hilbert schmidt norm. Check out this subsection on kernel embeddings of (in this case) distributions for example
Our work got in to web conf, WWW 2022. Big Q: will i end up in Lyon, France or will it go down as virtual? Happened to us for AAAI, CVPR, Globecom, FG so far like many of you in your own circles.
@ilyaraz2
+1 on the thread. countless stories. including mine: had H1b not picked in lottery twice post F1-and went back to home country. finally got h1b cap exempt in a univ. and then to an O-1 in a company before moving to a GC. many pillar to post last min uncertainties.was on J1 too.
Workshop on Statistics for the future of AI at
@mbzuai
Jan 2024. with keynote speakers including
@SusanMurphylab1
@XihongLin
, Tianxi Cai and Hongtu Zhu. in a month after their collaborative learning workshop in Dec! See you there (in person & virtual)
📢: The 87th FLOW talk is on Wednesday (30th November) at **5 pm UTC**.
Praneeth Vepakomma (MIT) will discuss "Recently engineered variants of split learning & theoretical advances towards ensuring its privacy."
Register to our mailing list:
Congrats to
@raskarmit
(advisor during PhD) on the Test-of-time award for Femto-Photography at Siggraph 2024! Trivia: Paper was in collab with Nobel laureate Moungi Bawendi.
Our exciting PoC paper on inducing differentially private communication for distributed ML via analog means with varying transmission powers:
@raskarmit
Rest in peace bro.. Young, jovial, fit marathoner, cricketer and biker..and a real good human being-most importantly. child hood & neighborhood frnd since my kindergarten. covid19...
Rest in peace brother. I've known you for more than 25 years. We made many memories. You were one of the reasons we had an amazing childhood. Such a jovial and sportive person you were. You will be missed.
@aminkarbasi
1.) The geometry of hypothesis testing over convex cones: Generalized likelihood tests and minimax radii Wei, Wainwright, Guntuboyina in Annals of Stat, 2019 & 2.) From Steiner Formulas for Cones to Concentration of Intrinsic Volumes McCoy, Joel Tropp in Discrete & Comp Geom.
Glad to share our collaborative tech report on Particle Flow Filters with 1 section done per author: a reminder that there are many different ways to work together!
April 9, Friday: Workshop on "Communication Efficient Distributed Learning" featuring amazing Martin Jaggi,
@AryanMokhtari
, Virginia Smith, Himanshu Tyagi
@hstyagi
,
@MountainOfMoon
. Please register (free) to attend and consider presenting a poster
@EugeneVinitsky
There are so many routes. its totally OK. Mine was extreme, i didnt even apply right away. ex: did ms in mathematical statistics, after CS ug & learnt more math (and math maturity) than ms in CS would ever have given. Then indust. 10 yrs & came to PhD program as I love research
teaser: looking fwd to making public our work based on Fourier Analysis of Boolean Functions..by mid of Jan. Happy New Year Everyone-hope you look back positively at 2022 when it's 2023. no pun :).
Hamdi Ulukaya (Chobani), speaker at Wharton MBA graduation ceremony: "It's great that you are a Wharton MBA, but please don't act like it". Don't let it get into the way in seeing people as people and all that they have to offer you regardless of their title or position.
Our notebook on "PLRV's", "Privacy Loss Distributions" & "Composition via Convolution" in differential privacy We empirically observe instances of theory in this Google's work as true: More in a blog ..
@raskarmit
@ArmaanBhojwani
@FrnkNlsn
The Wainright and Vershynin books have rightly become classics of the modern day indeed. Thanks for shedding light on the rest of them too-that I was unaware of..
@peter_richtarik
@NeurIPSConf
Same story here. I gave detailed reviews and engaged in discussion on all papers and a short review/comment to one (out of my expertise). Reg my 2 submissions-not one has discussed or commented on my detailed rebuttal and revision yet.
What is the distribution of Euclidean distance between two Gaussian samples? In the case when they are standard normal there is . What if the variance isn't 1 but the mean is still 0-how would the resulting distribution change?
Loved this version: Benjamini-Hochberg correction for multiple hypothesis testing, after having checked out a few different ones and reading a few materials before preparing to teach BH for my recitation class
@boazbaraktcs
This course by Boaz along with its sister course by Ankur Moitra have been phenomenal! Was really fun to scribe the lecture contents into the blog..
Join the summer school on "Trustworthy, safe, and governable AI" organized by
@adia_lab
in Granada, Spain:
A great lineup of expert speakers from academia and industry covering
#AI
,
#Economics
, and
#Computing
.
I always tell my students: if you only read paper published in the past five years, the probability that you will have any ground-breaking idea in your lifetime is nearly zero. The odds is probably less than winning a jackpot on a slot machine in Vegas...
April 9, Friday: Workshop on "Communication Efficient Distributed Learning" featuring amazing Martin Jaggi,
@AryanMokhtari
, Virginia Smith, Himanshu Tyagi
@hstyagi
,
@MountainOfMoon
. Please register (free) to attend and consider presenting a poster
here it is... We recognize and precisely study a new trade-off between privacy & welfare. prior work studies privacy-utility and privacy-fairness trade-offs so far..