It appears that I no longer have to assist any professor and neither do I have to associate with any. Indebted to my students, collaborators and friends. I remain open to internship offers.
Prof. Prakash Narayan, my information geometry teacher, gave me permission to share his lecture videos and handwritten notes for "Information Theoretic Methods in Learning". Here they are in dropbox:
Course description:
International students come to US for grad school and research as the center of academia is still in US because of them coming continually. In return, they are tortured with CPT, OPT and 100s of different bureaucratic university rules designed to make them suffer.
Nice advice. But as a STEM phd, I found something like this very difficult to follow. Should I not run my code that I just wrote? Should I stop thinking about a proof for my theorem because it is 5 pm? What if I am almost there? And for most of my PhD, I was almost there.
So you're going to do a PhD! Yay! Here's my best advice:
1) Don't work nights/weekends unless you have a hard deadline. Work 40hrs, no more. My most cherished grad school experiences were hanging out with my friends. Do not waste your youth providing unpaid labor. Don't do it!!
Happy New Year, Everyone! It is a good time to announce that I have just joined
@UCSanDiego
in
@HDSIUCSD
Multiple postdoc positions are available to work in information theory++, among other areas
Remember remember. They decline to do the same this year in the middle of 400k covid cases each day + massive shortage of oxygen and hospital/icu beds in India, because?
DEADLINE EXTENSION: The submission deadline has been pushed back 48 hours (now June 3rd, 23:59 UTC-12) to allow people affected by recent civil unrest and Covid19-related lockdowns some additional time to finalise their papers.
Prof. Alexander Barg taught these two courses in last two years 1) Modern discrete probability 2) Coding theory . Fantastic resources in the links, not to mention insight on how to put topics together. Someone had to share these.
Please dm/email me if you or someone you know is interested in doing a postdoc with a prompt start date at UC San Diego in the areas of information theory, coding theory, statistical learning or distributed optimization.
Haven’t been this happy teaching a class for a long time. I actually wish it went on for few more weeks - for the first time. Here are all the lecture notes for my essential coding theory for anyone interested
April 9, Friday: Workshop on "Communication Efficient Distributed Learning" featuring amazing Martin Jaggi, Aryan Mokhtari
@AryanMokhtari
, Virginia Smith, Himanshu Tyagi
@hstyagi
. Please register (free) to attend and consider presenting a poster
Found this picture of me in high school with the then captain of the Indian cricket team
@SGanguly99
I remember this day to be one of the happiest at that time - though the actual moment is a blur now
Someone, not too long ago, acknowledged me in their paper for giving them the main problem of the paper. Dude, I did not *give* you the problem to go away with. I lend you the problem with the expectation of you returning it to me with interest, i.e., the solution.
NeurIPS/ICML should desk reject 60-70% papers and recruit qualified reviewers to give proper review to the rest. Also desk rejections will save time for majority who would’ve gotten random decisions in the current model anyway.
UCSD Data Science has multiple Asst. Professors pos in all areas: theory and methods strengthening AI; NLP; Statistical Learning; Explainable AI; Games/Economic theories; Database/Data Security;Intelligent Systems; Optimal Control, Cybersecurity Systems.
I'd list top 10 of my achievements in 2021 but all of them are spending time with my family. If anything in my academic life I am most proud of, that's my three students graduating with their PhDs.
#BYE2021
The 2023 ITA Workshop (Information Theory and Applications W.) back to the Valentines Day week Feb 12-17, 2023 in San Diego. Stop everything and make your plans accordingly:
Arrived in Vienna for
@icmlconf
yesterday with an empty suitcase; after a trip to the industry booths, I am all set with t-shirts, caps, and fidget-spinners for the rest of my eurotrip. I suggest giving away shorts as swags so that I can come by Ryanair next time.
One would expect the parameter error of logistic regression to scale as \sqrt{d/n} cause well-known asymptotic normality (d=dim, n= samples). But in the special case of halfspaces, it scales as d/n. What causes this change & is it gradual? With
@djhsu
By far the most outstanding result of the past few years in coding theory. Kudos to Henry Pfister for taking up this ambitious project and delivering (again).
Tagging
@s_kudekar
and
@HamedSHassani
1st yr grad student Namiko Matsumoto and I have a paper (to appear in
@focs2022
) with a descriptive title "Binary Iterative Hard Thresholding (BIHT) Converges with Optimal Number of Measurements for 1-Bit Compressed Sensing (1bCS)". What're BIHT & 1bCS? 1/
I can't stop thinking about my hometown, Siliguri, in Darjeeling District, North Bengal, this New Year's Eve. In the foothills of the mountains, it was beautiful in the late 90s), and I could see Kanchenjunga, the 3rd highest peak of the world, from our roof
#HappyNewYear
My student and colleague
@soumyabrata_pal
defended his thesis successfully today. I have learned innumerable things from Soumyabrata in the last 5+ years, things that greatly influenced my research. I am really fortunate to know him as a researcher and as a person.
Just finished teaching a quarter-long undergraduate course on Algorithms for Data Science, all the lecture notes are available here for anyone to use: As an undergrad I didn't know most of these topics, but for my students these were quite alright.
One would expect the parameter error of logistic regression to scale as \sqrt{d/n} cause well-known asymptotic normality (d=dim, n= samples). But in the special case of halfspaces, it scales as d/n. What causes this change & is it gradual? With
@djhsu
When I applied for grad school from India, the hefty application fees meant that I could only apply to a small number of universities (8 as far as I recall), and had to do all kind of optimization- not all top schools, some in the middle, some safe bets etc. 1/2
@beenwrekt
One problem is Yoshua Bengio said this 4 years ago and then published 100 papers the next week: Since he is one of the most rewarded/celebrated ML researchers, we all understood what this means: not only publish more and more, but also be a hypocrite.
Being a wise man, he avoids being in this website. But that will not stop me from welcoming Dmitriy Drusvyatskiy as a new professor to
@HDSIUCSD
@UCSanDiego
here! This has been a great recruiting season
While we are hiring faculties, postdocs and new grad students in troves, should also focus on making the lives of existing students better. Why humiliate them with weird “rules” every six months when you can’t even pay them decent salaries.
Done teaching another installment of Algorithms for Data Science (designed for MS level)- oh the relief after submitting the letter grades
Anyway, here are the (annoyingly) annotated lecture notes for ne1 interested in major basic concepts that drive DS
Along with Raghu Meka and Rachel Ward, we are organizing this workshop at
@ipam_ucla
on computational/statistical gap of inference problems: Week of Feb 26, 2024 the one after ITA 2024 and ALT 2024 at San Diego
@EnCOREInstitut
One should just look at the background of our international students to understand that they will succeed in spite of every rule, every sanction, and every bias. Get your shit together, universities, and treat them with respect before it is too late.
Reed-Solomon Codes have been the central topic of coding theory - but they are very simple and all you need to know up to and including Sudan's list decoding can be summarized in just ~ 2 pages :-) So, in case anyone wants to teach RS codes:
Finally a news that’s been brewing for years.
@HDSIUCSD
and San Diego Supercomputer Center are coming together in forming the School of Computing, Information, and Data Science (SCI-DS), the 12th school in
@UCSanDiego
All credits go to
@GuptaUcsd
for making this happen.
What kind of papers go to ICLR? I was under the impression that its scope is less broad than NeurIPS and experiment-heavy. But now it seems exactly like NeurIPS/ICML. So it is just a third everything-goes conference?
We are looking for outstanding candidates for an assistant professor position in the areas of statistical foundations of data science at
@HDSIUCSD
@UCSanDiego
. Among areas of interest are the intersection of statistics and computation and related theory.
The highest rated paper in my NeurIPS bunch does something trivial - but the reviewers don't have much clue and/or benevolent (it is not even an ML paper imo). The lowest rated paper is also trivial - but reviewers are just mean.
Google is not the new Bell Labs. Even the comparison is laughable. MSR had a shot maybe 15 years ago, but they were too narrow- and by now a bottom line driven corporate wing like most other places. I don’t think Bell Labs is even possible in today’s environment.
@Noahpinion
No.
The research arm of Bell Labs was never about moonshots.
It was about hiring the best scientists into small departments (typically 5 to 15 people) and giving them resources and a *lot* of freedom to work on what *they* deemed most promising.
That's how you get breakthroughs.
Simons Institute's program on Theory of Reinforcement Learning kicks off next week:
Also in case anyone wants to check out the very interesting website about how RL works😉:
Anytime you cite a paywalled article instead of the arxiv version, you are lending legitimacy to for-profit closed source science-exploiters. Anytime you do anything, you lend legitimacy to the worst behavior under that model.
Suppose x is an n dim vector with only k nonzero entries. We can measure x by designing an m x n matrix A and looking at y=Ax. What should be the minimum m such that we can recover the support of x with an A that works for all x? This is a simple version of compressed sensing 1/5
All reviews/meta reviews should be public, so that there is some accountability. This is one of the reasons why even smaller scale theoretical conferences are cliquish and heavily biased towards some arbitrary topics. It is time to stop scratching each other’s back.
"Collusion Rings Threaten the Integrity of Computer Science Research" by
@MLittmanCS
@BrownCSDept
says attempts to "game" the peer-reviewed
#conference
publications system could undermine the mechanism and damage the ability to share
#research
effectively
@AndrewYNg
So what is your best estimate of the probability that if AI is not strongly regulated it will lead to human extinction in the next 30 years? If you are a true Bayesian you should be able to give a number. My current estimate is 0.1. I suspect Yann's is <0.01
All international conferences with deadlines in May should consider this. How’s it possible to submit papers in the current state if you are in India?
@NeurIPSConf
show you care for people outside of the US.
This is not an unreasonable request
@NeurIPSConf
. If the same thing happening in India was happening in the US rn, I would expect the same. here's a chance to show that you care about the global research community, not just where the centre of mass currently happens to be.
The ECE department of
@UCSanDiego
, is seeking faculty applications for all levels (below only asst. prof link) in the areas of information and communication theory, learning theory and signal processing. Please apply if interested
Everyone knows the difference between an arxiv preprint and a peer-reviewed Nature paper: the later will only be proven fraudulent after some press/publicity and promotion/professorship.
This is such an unobserved phenomenon in (engineering, cs, math, stat) academia that I am highly surprised that this is prevalent in these other fields. There is rampant elitism in our side - but that is more institutional (eg Stanford grad) and can be overcome w/ elite research
I have completed my PhD in half the allotted time whilst working, have 2 book contracts, 12 publications, 5 yrs HE teaching experience, & a charity that delivers high impact research. What do institutions see? A working class background. 0 interviews offered.
@PostdocVoice
If you are interested in an extended research visit to UC San Diego, along with a few colleagues, to work on a project related to foundations of AI and data science, please write to us. We will fund it and provide you the space. Sabbaticals welcome!
Deciding not to submit a paper to a conference sometime before the deadline is a great feeling. People (including me) often forget this. Ecstatic! Try it!
Mind-boggling number of people go through 4/6 years of college without having absolutely any idea what professors do. How do they think the things in their textbooks came to being?
Seeing data science being presented as something shallow in many popular articles/agendas is unsettling, because as far as I understand, inferring meaning out of data can be intellectually challenging and will require intricate tools from mathematics and other sciences.
Is it ok to hype up low quality bandwagon research from the position of extreme privilege to hog even more attention, and rip benefits of all publicity but to not be prepared to take criticism?
Twitter tells me
#NeurIPS2020
has a fine strategy. They can now reject a paper 3 times in a single cycle. Aug 7: Borderline Rejection with some hope, Sep 1: Rejection looming after discussion, Sep 25: Absolute Rejection 🥰 + 3 bad days.
Middle-schoolers are learning algorithms from
@B1ar2n3a
. Yesterday around this time she was submitting a paper, tomorrow she is set to give a seminar in Simons.
Wow! We are delighted & humbled to have received a generous, six-figure donation from
@jk_rowling
charitable trust, The Volant Charitable Trust () for our Covid Relief work in India. We are humbled by the support for our humanitarian work.
#COVIDIndia
once
@ylecun
told me (heavily paraphrased), it's not F=ma but \min (F-ma)^2. i didn't realize its importance, but it is perhaps the most enlightning perspective i've ever heard.
Fortunately, I had the exact same idea of a good day; and I found an advisor who believed this as well. Therefore, I had an enjoyable PhD experience. And I am still a bit sad that my PhD ended and still look forward to meeting my advisor all the time.
(2) A PhD is hard - you are trying to explore the unknown, and a large part of this will involve getting stuck, and trying things that don't work. Feedback is sparse. Don't measure success by papers. Every day you learned something is a good day.
Why are the blurry images of black holes interesting? Are humans supposed to stare at black holes? What scientific insight this visualization gives? Does any theory get validated by them? What physics do we know now that we did not before? Any good articles about this?
We are seeking postdocs in TILOS Institute (an NSF AI Institute ) at UC San Diego to conduct research on the foundations of AI and optimization, among other areas. Please apply here:
Two of my students have defended their thesis successfully this summer:
@rkmaityju
Raj Maity (next Microsoft) and Larkin Flodin (next Amazon). Congratulations 🎊 to both. I know both of them have to overcome many odds to be where they are now.
However with
@soumyabrata_pal
we are able to show that m~ k^1.5 is sufficient. There exists an A with that many rows that works universally and recovers a small superset of support. Check out this result scheduled to appear in the ITCS conference 5/5
Today at a group lunch a new grad student mistook the *only* female faculty attending the event as something other than a professor - even after she introduced herself as a professor. He started by assuming she is a grad student nonetheless 🤦♂️ How we keep getting these specimens
Unfortunately, like the best people of many fields, the worst practices of many fields are also in ML now. In engineering academia, everything is measured by paper count; and senior/respectable(?) people submit 100s of questionable papers to all kind of journals/conferences.
@suguna_misra
@MountainOfMoon
@lorisdanto
As a final comment, there are people who get 10-20 papers into subsequent NeurIPS/ICMLs. This is unheard of in theory. Max is probably about 5 papers, and definitely not back to back. Getting the requisite ideas is much harder and slower (vs ML where often ideas are fairly cheap)
Congratulations! Lázló Lovász and Avi Wigderson share the Abel Prize of 2021.
They receive the prize “for their foundational contributions to theoretical computer science."
#AbelPrize
#AbelPrize2021
TIL that ICLR is the
#1
conference in "Artificial Intelligence" according to Google Scholar Metrics () but it's still not included in . All rankings are silly and arbitrary, but this seems especially silly and especially arbitrary.