I'm in the magical Vienna
@iclr_conf
, presenting our work: at Halle B
#244
, Wed 8th, 10:45 — 12:45 & at The Google Booth:
#23
, Hall A, Fri 10th, 12:45 — 14:15.
Come, say heyo! Let's talk Trustworthy ML :)
#iclr
#iclr24
#ICLR2024
🪟 Quick glance at our recently accepted
#ICLR2024
@iclr_conf
work!
📜:
Want to learn from privacy enhancing aggregate data but still want to achieve robust & scalable near-instance wise performance?
Parity Checks from Coding Theory to the rescue!
(1/17)
Had my own fan-boy moment bumping into the Godfather of AI himself
@NeurIPSConf
Was star-struck & stuck trying to introduce myself before I could even get to explaining my research
But hey
@ylecun
if you find time, do visit my Spotlight and Oral papers :)
eight yrs ago i was a jr phd student over the moon to be sharing beers with the inventor of the lstm (
@HochreiterSepp
) at neurips in montreal w my first deep learning collaborator
@davekale
. hope you get to share a human moment w someone you admire this week.
Godfather 2, this time with a flavour of NLP
@iclr_conf
@chrmanning
was gracious enough to not refuse my offer of a selfie :)
Each conference I'm amazed at how humble my idols are 🫡
At the rate I run into celebrities, I should probably walk around LA next time 🤝
#ICLR2024
Had my own fan-boy moment bumping into the Godfather of AI himself
@NeurIPSConf
Was star-struck & stuck trying to introduce myself before I could even get to explaining my research
But hey
@ylecun
if you find time, do visit my Spotlight and Oral papers :)
Had a blast presenting our work to diverse audiences
@RealAAAI
, discussing creative questions, comments & suggestions y'day :)
Hopefully this prompts work overcoming bias even under tricky distribution shifts fostering
#TrustworthyML
I'm around to chat more!
#AAAI2024
#AAAI24
Heyo Vancouver! 🇨🇦
I'm
@RealAAAI
, presenting our work at the Poster Session on the 23rd, Friday at 7 pm! See you there & let's talk
#TrustworthyML
!
Feel free to hit me up wherever!
Looking forward to expanding my research horizons & exploring this week :)
#AAAI24
#AAAI2024
Hello folks!
Bubbling with excitement to be heading over to New Orleans for my first in person conference at
#NeurIPS2023
!
Would be presenting two innovative approaches to solving important problems of today and tomorrow representing
@GoogleAI
@GoogleIndia
Details below
(1/7)
🚨 Fresh
#AAAI24
@RealAAAI
Paper Alert
@GoogleAI
@Google
Paper:
Code:
It has been observed that algorithms that optimize predictive accuracy that are fed pre-existing biases further learn & then propagate the same biases.
(1/12)
🍩 Celebrating the acceptance of "Learning from Label Proportions: Bootstrapping Supervised Learners via Belief Propagation"
@iclr_conf
, with a chocolatey box of happiness courtesy
@dunkindonuts
!
📃:
Details to follow soon! 🇦🇹
#ICLR2024
@GoogleAI
@Google
🪟 Quick glance at our recently accepted
#ICLR2024
@iclr_conf
work!
📜:
Want to learn from privacy enhancing aggregate data but still want to achieve robust & scalable near-instance wise performance?
Parity Checks from Coding Theory to the rescue!
(1/17)
🍩 Celebrating the acceptance of "Learning from Label Proportions: Bootstrapping Supervised Learners via Belief Propagation"
@iclr_conf
, with a chocolatey box of happiness courtesy
@dunkindonuts
!
📃:
Details to follow soon! 🇦🇹
#ICLR2024
@GoogleAI
@Google
Fanboy moment 2.0, the ever so humble & ever so cheerful
@zacharylipton
patiently explaining why true trustworthy ML is not like a condiment 🍝
I hope someday someone takes a similar photo with me
@NeurIPSConf
, with me on the other side of the equation :)
#NeurIPS2023
#NeurIPS23
eight yrs ago i was a jr phd student over the moon to be sharing beers with the inventor of the lstm (
@HochreiterSepp
) at neurips in montreal w my first deep learning collaborator
@davekale
. hope you get to share a human moment w someone you admire this week.
I'm in the magical Vienna
@iclr_conf
, presenting our work: at Halle B
#244
, Wed 8th, 10:45 — 12:45 & at The Google Booth:
#23
, Hall A, Fri 10th, 12:45 — 14:15.
Come, say heyo! Let's talk Trustworthy ML :)
#iclr
#iclr24
#ICLR2024
Got dehydrated & famished, but in the end, Day 1 of Workshops was all worth it, so many amazing folks doing such amazing work! Talked fairness or/& distribution shifts across 2 unnecessarily distanced rooms.
@NeurIPSConf
please keep similar workshops close by next time, thanks :)
Come check out our spotlight work at Room 252 - Room 254 (Level 2) and let's talk more about fairness and distribution shifts :)
Amazing conversations at the morning poster session
@afciworkshop
See you at the talk at 13:00 and the afternoon poster at 16:50 to 1730!
#NeurIPS2023
Heyo Vancouver! 🇨🇦
I'm
@RealAAAI
, presenting our work at the Poster Session on the 23rd, Friday at 7 pm! See you there & let's talk
#TrustworthyML
!
Feel free to hit me up wherever!
Looking forward to expanding my research horizons & exploring this week :)
#AAAI24
#AAAI2024
🚨 Fresh
#AAAI24
@RealAAAI
Paper Alert
@GoogleAI
@Google
Paper:
Code:
It has been observed that algorithms that optimize predictive accuracy that are fed pre-existing biases further learn & then propagate the same biases.
(1/12)
Oh how I wish I was applying for grad school 5 years earlier!
Also encouraging to see folks from all stages of life contribute to this growing field; but now it just adds to the pressure of having objective metrics (papers/citations/now this) define your research.
Not ideal.
2019: Need NeurIPS papers to get postdoc/faculty position.
2024: Need NeurIPS papers to get PhD student position.
2029: Need NeurIPS papers to get into undergrad program.
2034: Admission to NeurIPS onsite daycare requires accepted paper.
Stop by the
#ICLR2024
Google booth today @ 12:45 PM to hear
@_toolazyto_
discuss how to learn from privacy-enhancing aggregate data to achieve robust, scalable performance using the secret ingredient: The Belief Propagation Setup inspired by parity checks.
Come check out our spotlight work at Room 252 - Room 254 (Level 2) and let's talk more about fairness and distribution shifts :)
Amazing conversations at the morning poster session
@afciworkshop
See you at the talk at 13:00 and the afternoon poster at 16:50 to 1730!
#NeurIPS2023
Improving Fairness-Accuracy tradeoff with a few Test Samples under Covariate Shift: (Spotlight)
A novel solution to ensure fairness under distribution shift, a Spotlight @ Algorithmic Fairness Workshop
@afciworkshop
@ 9:00 CST Fri 15 Dec.
(3/7)
Wait so all this while us Indians were pronouncing it correctly? And the Americans were not?
(Ton-Don not Tan-Don for those curious)
Need to go down school name origins rabbit hole :)
If anyone wants to chat about Classical Music, Football, Reading, Randomest Trivia, Touring Reccos, anything under and beyond the sun, I'm more than happy to partake in that :)
And if anyone wants arguments for why more conferences should be held in Europe, PFA :))
#ICLR2024
I'm in the magical Vienna
@iclr_conf
, presenting our work: at Halle B
#244
, Wed 8th, 10:45 — 12:45 & at The Google Booth:
#23
, Hall A, Fri 10th, 12:45 — 14:15.
Come, say heyo! Let's talk Trustworthy ML :)
#iclr
#iclr24
#ICLR2024
Really really cool and well presented work by a dear friend on a topic that is the need of the hour :)
Also one of the brightest and richest minds on the grad school application market, gobble him up while he's available 😤
Extending an LLM for new knowledge sources is tedious—fine-tuning is expensive/causes forgetting, LoRA is restrictive.
Excited to share our work where we show that an LLM can be efficiently *composed* with specialized (L)LMs to enable new tasks!
🧵(1/8)
And that's a formal (tried a pun here, pls appreciate) wrap,
@NeurIPSConf
was so hectic & so happening, yet so so fun. Seeing your academic heroes irl, and them patiently and humbly listening to you ramble on, was my highlight of
#NeurIPS2023
What was yours?
#NeurIPS
#neurips23
Presenting this work as an Oral at the Regulatable ML Workshop in 20 mins at Room 215-216, come see me wearing my blazer for the first time since my college farewell! 😤
Thanks for the engaging poster discussions, we'll be around till the evening, come and say hi!
#NeurIPS2023
Improving Fairness-Accuracy tradeoff with a few Test Samples under Covariate Shift: (Spotlight)
A novel solution to ensure fairness under distribution shift, a Spotlight @ Algorithmic Fairness Workshop
@afciworkshop
@ 9:00 CST Fri 15 Dec.
(3/7)
Research Week with Google - officially a wrap!
Extremely energising to be with students and see their research curiosity! Till next time!
Special thanks to our amazing speakers, ACs, organisers and Program Chairs !
I feel the majority of people hold popular opinions just to please the majority. Most people I know have never tasted a pineapple pizza but hate it with all their heart.
It's concerning when you think of this phenomenon's repercussions in major decisive opinions.
for the past 2 weeks I’ve been reading every abstract from
#neurips2023
here are my notes about what I discovered:
i think everyone can learn something new from this, and I hope this resource is useful!
The enthusiasm and excitement for "science" screenings and the landing eclipsed that of several movies.
Hope this is the beginning of great strides in scientific and technological progress in and for our country 🇮🇳
#Chandrayaan3
#Chandrayaan3Landing
#ISRO
This is probably my most favourite random sports fun fact.
And my proudest moment as a quiz master is self-discovering this, and making a workoutable question.
Seeing the Eurekaaaaa moment emerge on the faces of the teams with oohs and aahs, the best part of QMing period.
🤌
India Wants To Know: India's First Panel Quiz Show
The way you win football games is by scoring goals.
To win a football game you score goals, usually more goals than the opposition. And ideally — and this is the important bit — into the other team’s net.
But Barbados football team had a strategy to win by scoring own goals. A
RGD, a new variant of the stochastic gradient descent algorithm, re-weights data points during each optimization step based on their difficulty. Learn how this reweighting algorithm can boost performance on unseen data across various learning tasks →
The KQA w/
@DrunklingP
is hosting Ed.
#32
of its fortnightly pub quiz on Wednesday!
830pm on 4th October at Drunklings Sizzlers and Pub, Koramangala. QM: Shreyas Havaldar. Quiz for teams of 2, open to all!
Buy your team's ticket at
Learning from Label Proportions: Bootstrapping Supervised Learners via Belief Propagation: (Oral)
A coding theory inspired take on the learning from aggregate data problem, an Oral at the Regulatable ML Workshop at 10:15, Sat 16th Dec CST.
(2/7)
This has increasingly become a problem for aspiring people from the "Global South", and what's even more concerning is the fact that the difficulties independent researchers might be facing would be even worse.
Need to act actively on making conferences more accessible for all.
@culturaltutor
One of the most intriguing things is, across cultures spanning different continents, the astronomical object associated with the day is also the same. For eg., in India, Sunday corresponds to "Ravivar", which translates quite literally to Sunday. Likewise for other days.
How so?
@Trevornoah
Unfortunate that people traveled hours in abnormally high traffic; spent loads commuting, only to reach a wedding hall with misleading seat layout. Upsetting for us who waited eagerly for months. Kudos to
@Trevornoah
for bravely not doing the disappointing show. Such a heartbreak
Also exhilarated to announce that "Improving Fairness-Accuracy tradeoff with a few Test Samples under Covariate Shift" has also been accepted to
#AAAI2024
!
Thrilled to present the work at two prestigious venues discussing fairness & robustness with the best minds around!
(4/7)
Am I the only one who remembers years, in birthday to birthday chunks? Also helps that my birthday is near the academic calendar, but just makes sense to start mentally counting years as whole from the birthday; cause like otherwise there would be a random halfish year somewhere
@AkshayaS90
@bertyashley
Sort of? Origins of nicknames does have an etymology flavour to it, and more importantly all people were famous enough for me to know them, so good TIL day :)
One of the best informative YT channels around. Much deserved!
Beautiful animations and simple & elegant explanations behind some of the most intriguing concepts in mathematics. And the invariable BAMs to keep you in high spirits!
Check out the channel if you haven't already!
Hooray!!! My StatQuest youtube channel crossed the 900,000 subscriber milestone over the weekend! TRIPLE BAM!!! To celebrate, here are my top 8 videos with over 1 million views! 🧵 1/9
Saw this posted on LinkedIn as a leader’s recommended reading list (or something along those lines).
Once again I am begging of you, please read more fiction. I promise that it will do more for you vs another productivity optimization book.
It would be interesting to see how industrial and academic research directions pivot, but the development of safe, robust and interpretable models is definitely the need of the hour!
Great tldr to glance at the relevant salient points, thanks
@kchonyc
!
a number of weird definitions and weirdly specific points, but overall, worth reading it to see which areas are considered as priorities by WH. in this 🧵, let me copy-paste those few weird/interesting/specific points i found reading it.
It wouldn’t be an exaggeration to say that Aug 15, 1992 changed Indian cinema forever.
This was the day when Mani Ratnam’s Roja was released.This also marked the debut of the musical storm from Madras.
Sit back, relax, grab coffee & headphones.
A thread🧵on
@arrahman
's music!
Presenting this work as an Oral at the Regulatable ML Workshop in 20 mins at Room 215-216, come see me wearing my blazer for the first time since my college farewell! 😤
Thanks for the engaging poster discussions, we'll be around till the evening, come and say hi!
#NeurIPS2023
Learning from Label Proportions: Bootstrapping Supervised Learners via Belief Propagation: (Oral)
A coding theory inspired take on the learning from aggregate data problem, an Oral at the Regulatable ML Workshop at 10:15, Sat 16th Dec CST.
(2/7)
@LongFormMath
This calls for a book recommendation!
is a nice way to discover his other quirks & experience his journey, while still having a nice ELI5 backdrop of the field of logic & discrete math.
The author list is illustrious to say the least :)
Honoured to have been a part of this bright and growing group of people, an amazing place with amazing people all around, cannot overstate the enriching nature of edification :)
Highly recommended for people at various stages of undertaking and providing education.
We bring the brochure of CSE IITH containing details of our journey so far, faculty profile, various teaching and research activities, achievements of our students and faculty, projects undertaken, and impact to the society.
The brochure is available at:
Another common challenge that models deployed in real world situations face is that of Covariate Shift where the distribution of feature vectors across training and testing changes, however the optimal label predictor conditioned on input remains the same.
(3/12)
Also utilizing this thread to plug in an advert for myself, I'd be applying to Grad School (Research Positions) this cycle (Fall '24) and would love to connect with like-minded folks, labs and advisors on Privacy, Fairness, Interpretability, Robustness, Explainability!
(5/7)
With machine learning systems becoming ubiquitous & universal, ensuring fairness in systems that are used to make important decisions, having tremendous societal impact, like in criminal justice, loan approvals, has become increasingly important.
(2/12)
@sumanthd17
@Google
@GoogleAI
It was great fun having you around :)
Lemme know when you're done with the book I'll steal it again :p
That's another excuse for you to come to Bangalore 😤
Aise hi phodte rehna 🫶
Should we bow?
Yes, he's the king.
So so heartening to see chess amass the public following it fully deserves. Hope like Vishy's story, this goes on to inspire people to take up the sport and create many such stories.
The future of chess is India ❌
The now of chess is India ✔️
17-year-old Grandmaster D Gukesh is the Champion of FIDE Candidates Chess Tournament 2024!!
Gukesh scored 9/14 points to clinch the first place. A huge moment for Indian Chess! After Vishy Anand, he is only the second Indian player to win the Candidates Tournament.
With this
Joint work w/ fascinating collaborators & mentors comprising
@chauhanjatin100
, Karthikeyan Shanmugam,
@jayjaynandy
& Aravindan Raghuveer.
This work was completed at my Pre-Doctoral tenure at Google Research. Indebted to
@ManishGuptaMG1
&
@divy93t
for the opportunity.
(12/12)
🚨 Excited to share our
#ICML2023
work on the Feature Sieve, by which we automatically identify and suppress irrelevant or spurious features in deep networks, hence mitigating simplicity bias.
Paper:
References:
[1] Zhang, Brian Hu, Blake Lemoine, and Margaret Mitchell. "Mitigating unwanted biases with adversarial learning." Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society. 2018.
#fairness
#trustworthyml
#robustness
Towards improving fairness under this highly challenging yet realistic scenario, we make 3 contributions.
1⃣ We propose a novel composite weighted entropy based objective for prediction accuracy which is optimized along with a representation matching loss for fairness.
(6/12)
Do hit me up on Twitter(X)/Whova/email, or any platform you prefer if any of these ideals inspire your work!
Would be delighted to just hang-out and discuss AI/ML or even anything under or beyond the sun. Hyped to meet amazing people doing amazing work & explore NOLA!
(6/7)
Okay, because this tweet (post?) is getting so many views I ought to plug my work here:
Please DM or come say heyo to me!
Always up for a chat, on research and beyond :)
Hello folks!
Bubbling with excitement to be heading over to New Orleans for my first in person conference at
#NeurIPS2023
!
Would be presenting two innovative approaches to solving important problems of today and tomorrow representing
@GoogleAI
@GoogleIndia
Details below
(1/7)
After 3 virtual yrs, we are excited to host 'Research Week with Google' in Bangalore. In-person keynote from
@JeffDean
and more awesome research discussions.
Undergrad / MS / PhD student in India? Apply before Nov 28. Pls spread widely!
@ManishGuptaMG1
You can learn a lot about history just by looking at the words we use.
Like algorithm, which is descended from the name of a 9th century Persian polymath called al-Khwarizmi.
So, from romantic to cynical, here are the strange stories behind 12 incredibly normal words...
More people boarding in an already jam packed train, even the AC is not working.
It has become utterly suffocating and people are unable to breathe.
No respite or any official announcement at all.
I humbly request someone to take cognizance of the issue.
@tiwarishabh16
@UCBerkeley
Big Congrats
@tiwarishabh16
! Berkeley is fortunate to bag you by winning against a lot of suitors; they better treat you well :p
All jokes apart, massive achievement, very well deserved, onwards and upwards!
Most SOTA fairness methods like Adversarial Debiasing's [1] fairness & accuracy go for a toss as we move from a no covariate shift to high covariate shift regime. Distributionally Robust & Importance Sampling based methods collapse as well!
So what can we do differently?
(5/12)
@boknilev
It actually might be interesting to look at D=P(A|Y=1) - P(A|Y=0) per attribute, even if data is skewed, but the predictor is "fair" wrt a sensitive attr A, D~0. But if the predictor was indeed "unfair", it might be the case that for the underprivileged A, D is significantly -ve.
"When life gives you a lemon, wing it right back and add some lemons of your own!"
Gem of a comic strip that has always been a part of my life via room posters, books and even my yearbook (which never came out) quote :)
Timeless and Ageless classic, rightfully ended on a high!
“Calvin & Hobbes” might be the purest comic strip ever in the artistic sense.
Despite its success — syndicated in 2k+ papers and 45m books sold — creator Bill Watterson avoided licensing.
By not licensing the IP for games and toys, he left an estimated $400m on the table.
He
@fooobar
Aapka mere lea doguna pyaar aakhir dikh hi gaya :) Jokes apart, thanks for being such a frank, humble & homely presence. Hope you keep inspiring people, as you no doubt did all of us. One day I'll take 5 points off your serve 😤 Karo har maidan fateh hamare Bharat ka Gaurav 🇮🇳🫶
Therefore the model may make wrong predictions when deployed or more seriously can slowly degrade over time when the shift is gradual.
The problem that we study in this paper is at the juncture of the above two hard problems: ensuring fairness under covariate shift.
(4/12)
This is the kind of science content I'm on Twitter for.
This is the kind of paper announcement we all researchers should do.
Such a beautiful ELI5 way to start simple, catch the curiosity , bring in the state of affairs, introduce your work and talk about the future!
🫡
@fooobar
Trying to unnecessarily stick to 4 seamers "safe-choice" narrative in England. No. 1 Test Bowler against a team with 4 left-handers is such a no brainer. And it's not like he can't bat either, waiting for the retrospective analysis where the team would accept their mistake :(
Was on my knees, breathless and about to puke by the time Montiel walked up to take his shot.
In the end, this match, which gave me (and a billion others) a hundred heart attacks was worth all of that pain.
CHAMPIOOOOOOOOOONS!
@rajnish_rg74
Yes that I've tried and am not the biggest fan of, would like to thank our first year LitR/OP party at Onesta for these Pizza discoveries 👍
1⃣ On the kNN graph induced by the covariate embeddings, perform belief propagation (BP) to obtain pseudo-labels that respect edge & bag constraints.
2⃣ Fit an MLP to the instance pseudo-labels & bag aggregate label.
Embedding learnt used to further refine the kNN graph.
(6/17)
Several exciting future directions to ponder upon, including exploring alternate energy potentials & understanding the convergence at such a scale!
Would love to chat more & brainstorm intriguing ideas towards Private & Trustworthy ML!
Looking forward to Vienna 🇦🇹 :)
(16/17)