Amartya Sanyal
@AmartyaSanyal
Followers
2,009
Following
1,236
Media
116
Statuses
1,590
Assistant Professor @DIKU_Institut @UCPH_Research , Ex Postdoc @MPI_IS @ETH_ AI Center @CompSciOxford , @StHughsCollege , @turinginst , @cseatiitk
Tübingen, Deutschland
Joined June 2014
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Columbus
• 402353 Tweets
Bills
• 140824 Tweets
無料クーポン
• 93411 Tweets
Bret
• 91895 Tweets
Jets
• 90337 Tweets
Yankees
• 73877 Tweets
連休明け
• 72383 Tweets
#WWERaw
• 59039 Tweets
Geraldo
• 44264 Tweets
Rodgers
• 42574 Tweets
Rogan
• 32240 Tweets
Boulos
• 31272 Tweets
Soto
• 29734 Tweets
緊急入院
• 29021 Tweets
当選確率2倍
• 26631 Tweets
Hail Mary
• 25612 Tweets
Buffalo
• 20321 Tweets
Josh Allen
• 19174 Tweets
#BUFvsNYJ
• 15514 Tweets
#RepBX
• 13542 Tweets
Stanton
• 10390 Tweets
@francoisfleuret
@Biz_Ukraine_Mag
Evacuation route out of Mariupol was mined, Red Cross says via
@BBCNews
21
417
801
I am not sure what others think but my personal opinion is
You don't need to have a "clear and strong interest in a specific research problem" to start a PhD, that's something you develop while doing your PhD.
If you don't have a clear and strong interest in pursuing a specific research problem for the next 4-6 years of your life, don't do a PhD. Go work for one of the many cool AI startups or research labs in industry: similar profile, faster pace, better pay, shorter commitment.
3/4
7
4
172
16
14
293
📢 Job update: Thrilled to announce I will be joining as Assistant Professor in Machine Learning at University of Copenhagen
@DIKU_Institut
next summer.
I will be taking PhD students (stay tuned for ads in a few months). Till then you can find me postdoc-ing at
@MPI_IS
.
29
9
252
I'm looking for PhD students to work with me on the topics of Privacy, Robustness, Unlearning, and general learning theory.
Apply and reach out to me if interested!
Also looking for postdocs, advertisement will be out shortly. Reach out if interested 🙂
The
#ELLISPhD
application portal is now open! Apply to top
#AI
labs & supervisors in Europe with a single application, and choose from different areas & tracks.
The call for applications:
Deadline: 15 November 2024
#PhD
#PhDProgram
#MachineLearning
#ML
1
101
278
4
70
241
We have at least one PhD opening (with me) in the domain of Differential Privacy/Robustness in ML.
(with many amazing collaborators/co-advisors inside/outside Copenhagen)
Start: Sep, 24
Location: University of Copenhagen
@DIKU_Institut
@AiCentreDK
.
See
The portal is open: Our
#ELLISPhD
Program is now accepting applications! Apply by November 15 to work with leading
#AI
labs across Europe and choose your advisors among 200 top
#machinelearning
researchers!
#JoinELLISforEurope
#PhD
#PhDProgram
#ML
6
175
417
0
35
116
I'll be attending
#ICML2024
in Vienna
& present our work on privacy, collective action, OOD robustness, and safety in large models.
Ping me if you want to chat :)
🚨 I'm also hiring a PhD and a Postdoc to start next year. If you think we will be a good match, DM/email me ;)
0
7
107
Very happy to say that I passed my D.Phil viva with minor corrections.
Much thanks to
@ilssontfous
and Phil Torr for tolerating me for the last 3 and a half years.
And Pranjal Awashthi and Pawan Kumar for examining my defence.
Oxford has, undoubtedly, been great!
9
0
102
Recently joined
@ETH_AI_Center
in
@CSatETH
&
@MPI_IS
as a postdoctoral fellow where I will work with Fanny Yang and Bernhard Schölkopf on theoretical and empirical topics in robustness, fairness and privacy in machine learning.
Let me know if you are keen to collaborate :)
4
1
97
Delighted to say that this paper was rejected from NeurIPS2020.
Neural Networks can "benignly overfit" some label noise. But is it truly benign ?
Turns out it provably hurts Adversarial Robustness
📢📢Paper: How Benign is Benign Overfitting ? 📢📢
New work with
@puneetdokania
, Varun Kanade (
@ilssontfous
) , Phil Torr
@OxfordTVG
Thread 👇🏾
2
15
72
1
0
77
Neural Networks can "benignly overfit" some label noise. But is it truly benign ?
Turns out it provably hurts Adversarial Robustness
📢📢Paper: How Benign is Benign Overfitting ? 📢📢
New work with
@puneetdokania
, Varun Kanade (
@ilssontfous
) , Phil Torr
@OxfordTVG
Thread 👇🏾
2
15
72
Very lucky and happy that we managed to have both our papers accepted at
#ICLR2021
1. Spotlight - How Benign is Benign Overfitting w/
@ilssontfous
2. Led by
@paudejorge
on Pruning at Initialization w/
@BehlHarkirat
Co-authors -
@puneetdokania
@OxfordTVG
Details below
1/4
6
11
71
#NeurIPS2022
Very happy to receive the Top Reviewer Award for NeurIPS 2022.
I hope my reviews were helpful for the authors and the wider community who will read the papers.
1
1
64
The call for this position is now public.
This will be jointly supervised with Prof. Amir Yehudayoff at
@DIKU_Institut
, Prof. Varun Kanade
@CompSciOxford
(including a stay in Oxford for 3-6 months), and myself.
RTs appreciated
I will soon be hiring another PhD student to work with me at
@DIKU_Institut
@MLSectionUCPH
in topics related to
1⃣ Trustworthy ML (Differential privacy, Robustness) and
2⃣ Leveraging unlabelled data in ML
Reach out to me via DM/email if you are interested and spread the word!
0
10
53
2
26
55
My DPhil (PhD) thesis is now online for public access. 😀
0
5
58
Very excited to be awarded my first ever grant (from
@HaslerStiftung
) to work on "Understanding the Interaction of Privacy and Fairness in modern ML Algorithms".
I am looking to host inters at PhD/masters level @
@ETH_AI_Center
@CSatETH
this fall. DM me if you're interested!
2
12
58
Very excited to have just started an internship with
@egrefen
at
@facebookai
.
Feels a bit weird to be not meeting all the amazing people in office physically but I'm extremely lucky to be holed up in the beautiful St Hugh's during this time.
3
3
56
Happening Now
Come join the live Oral session 1 for our Spotlight paper in ICLR 2021 if you want to ask me questions about our paper
"How benign is benign overfitting?"
Neural Networks can "benignly overfit" some label noise. But is it truly benign ?
Turns out it provably hurts Adversarial Robustness
📢📢Paper: How Benign is Benign Overfitting ? 📢📢
New work with
@puneetdokania
, Varun Kanade (
@ilssontfous
) , Phil Torr
@OxfordTVG
Thread 👇🏾
2
15
72
0
14
54
Very excited to be co-organising the 3rd
@satml_conf
in sunny Copenhagen next summer ⛱️.
This is arguably the best conference to submit your work in the intersection of trustworthy ML and security.
@MLSectionUCPH
@DIKU_Institut
@AiCentreDK
🎉Exciting news! The 3rd
@satml_conf
will be held in Copenhagen 🇩🇰 from April 9-11, 2025. We’re thrilled to have
@AmartyaSanyal
as our General Chair.
📢 More info on the SaTML website:
Stay tuned for updates on our Call for Papers and Competitions.
0
32
77
0
9
53
Today is my last day as
@ETH_AI_Center
postdoctoral fellow in
@CSatETH
@ETH
.
It has been a rewarding experience to learn from
@FannyYangETH
and other colleagues in the lab including
@alexandrutifrea
@DonhauserKonst
.
Exciting times lie ahead!
1
1
53
I will soon be hiring another PhD student to work with me at
@DIKU_Institut
@MLSectionUCPH
in topics related to
1⃣ Trustworthy ML (Differential privacy, Robustness) and
2⃣ Leveraging unlabelled data in ML
Reach out to me via DM/email if you are interested and spread the word!
0
10
53
Excited to be at
#Paris
for
@RiceUniversity
's Applied Algorithms for ML Workshop with such an exciting program.
P.S. Also looking forward to giving my own talk
0
2
49
Excited about this new preprint with Daniil and
@Kristo
We show a Ω(log T) lower bound for Differentially Private Online learning, even for finite Littlestone classes. This shows a separation between DP and non-DP online learning in mistake bound model.
1
7
47
Very happy to join
@MPI_IS
as a postdoc with and humbled to learn about causality and more from the amazing colleagues here.
If you're working here or just passing by and want to chat (or show me what the good spots in the town are) ping me!
Today is my last day as
@ETH_AI_Center
postdoctoral fellow in
@CSatETH
@ETH
.
It has been a rewarding experience to learn from
@FannyYangETH
and other colleagues in the lab including
@alexandrutifrea
@DonhauserKonst
.
Exciting times lie ahead!
1
1
53
0
0
41
Very happy that this paper was accepted to
#COLT2024
on the separation between private and non-private online learnability. Joint work with
@dmitrievdaniil7
and Kristof Szabo.
There are still few open problems that remain to fully settle this question :)
Excited about this new preprint with Daniil and
@Kristo
We show a Ω(log T) lower bound for Differentially Private Online learning, even for finite Littlestone classes. This shows a separation between DP and non-DP online learning in mistake bound model.
1
7
47
0
5
40
📢Worried about📢
1⃣ Terrible accuracies OR high computational cost with Differential Privacy ?
2⃣ Loose privacy parameters ε > 1 ?
3⃣ Unfair pre-training on datasets similar to private dataset ?
Presenting a new semi-private algorithm
🚀🚀PILLAR🚀🚀
1
11
37
Couldn't be more delighted and humbled to be giving a talk tomorrow on this paper at my undergraduate university - CSE department of IIT Kanpur.
There is hardly any institution I owe more than what I owe IIT Kanpur for whatever little achievements I have today.
@cseiitk
#IITK
Neural Networks can "benignly overfit" some label noise. But is it truly benign ?
Turns out it provably hurts Adversarial Robustness
📢📢Paper: How Benign is Benign Overfitting ? 📢📢
New work with
@puneetdokania
, Varun Kanade (
@ilssontfous
) , Phil Torr
@OxfordTVG
Thread 👇🏾
2
15
72
0
4
38
Deadline to apply for this PhD position is coming up in a bit over two weeks!
Apply if you want to work on theoretical and/or applied topics on trustworthy ML with an amazing couple of researchers and yours truly.
Email me for any questions :)
The call for this position is now public.
This will be jointly supervised with Prof. Amir Yehudayoff at
@DIKU_Institut
, Prof. Varun Kanade
@CompSciOxford
(including a stay in Oxford for 3-6 months), and myself.
RTs appreciated
2
26
55
0
5
37
Two of our papers on Differential Privacy --
1⃣ PILLAR, a semi-private learning algorithm, was accepted to
@satml_conf
2⃣ Our work on DP data release was accepted to
@aistats_conf
👇
Joint work with
@FannyYangETH
@francescopinto
@DonhauserKonst
@FraPintoML
Yaxi Hu and others.
1
6
37
#COLT2024
was amazing, as all other COLTs. Thanks to all organisers for the very smooth organisation and Edmonton for the many sunny days.
Looking forward to next year's COLT at Lyon.
0
1
36
In coming weeks, I'll be ✈️ a bit. I'll present posters and give talks on our recent works on robustness/privacy/fairness in ML. If you're interested I'd love to chat (email/DM me)
Dates below 👇
21 - 25 Sep Helsinki
26 - 29 Sep Boston
30 Sep - 3 Oct NYC
4 - 5 Oct Ann Arbor
0
0
34
Excited to be organising the
@TMLunLimited
at
@iclr_conf
in just one week. You can find
@alexandrutifrea
@jainprateek_
@sarahookr
in person in Kigali!
Our excellent invited and contributed talk list is now online.
1
6
33
Last two arxived papers 🚀 this year:
1. On privacy, fairness, and long-tailed data
How unfair is private learning ?
""
2. On certified robustness in practice
"Certified defences hurt generalisation"
Longer threads in the new year
1
2
30
Very excited to visit Toronto next week after nearly 10 years to attend
@satml_conf
.
If you are interested in the PhD position below 👇and will be in Toronto next week, please reach out :)
The call for this position is now public.
This will be jointly supervised with Prof. Amir Yehudayoff at
@DIKU_Institut
, Prof. Varun Kanade
@CompSciOxford
(including a stay in Oxford for 3-6 months), and myself.
RTs appreciated
2
26
55
0
3
30
First of our two papers accepted to
@icmlconf
this year is 👇work led by
@yaxi_hu
on quantifying the loss of privacy due to non-private pre-processing in a DP pipeline.
See you in
#ICML2024
1/2
Our work with
@AmartyaSanyal
and
@bschoelkopf
on privacy of pre-processing is online.
Exciting conclusion: Non-private data-dependent pre-processing can be integrated into a DP ML pipeline with minimal degradation of privacy.
Paper:
Visualisation 👇
0
1
16
3
2
30
1/ BLOG ALERT!! Classical ideas like dimensionality reduction can be successfully exploited in deep learning by encouraging the feature representations to lie in a low rank affine subspace.
Blog:
Paper:
@puneetdokania
@OxfordTVG
1
6
28
Today was my first lecture at
@CSatETH
with my own material on Computational learning theory and we ask feedback after lectures to clear the muddy points.
This surely fills me with joy and takes away the impending doom of ICLR reviews.
0
0
29
I am going to be in New Orleans for NeurIPS! Please DM if you want to chat.
I am also in the job market for tenure-track positions. Please reach out if you think I might be a good fit.
What I have been doing recently is partially reflected on my website
0
3
28
Understanding properties of neural networks by intervening on the datasets can lead to understanding of complex phenomena like catastrophic over fitting.
I'll not be at
#ICML2023
but if interested, reach out to
@gortizji
who said he can be found 24x7 on the beaches. ⛱️
Catastrophic overfitting can be induced with discriminative non-robust features
Guillermo Ortiz-Jimenez, Pau de Jorge, Amartya Sanyal et al..
Action editor: Jakub Tomczak.
#adversarial
#overfitting
#robust
0
3
22
0
0
27
Memorization need not be at odds with learning!
Are neural networks learning or memorizing? It must be learning, otherwise how they generalize so well. But hey, maybe memorization is not against generalization, and maybe necessary when there's little training data for some classes. by
@vitalyFM
& Zhang.
4
46
291
1
4
27
What a nice way to express what can be hoped from ML theory ?
By David Haussler
2
4
27
🚀🚀Our
@iclr_conf
workshop on Pitfalls of limited data and computation for Trustworthy ML
@TMLunLimited
is now accepting submissions on
@openreviewnet
till ⏰⏰Feb 8, 2023 ⏰⏰
We have an amazing PC as well as list of invited speakers ()
Happy new year everyone!
We are excited to announce that will be accepting submissions starting from 🚀Jan 15, 2023🚀 on
⏰ Deadline February 8th, 2023, 11:59 PM AoE
See the full CfP on our website
0
2
3
0
6
26
We are presenting at 5 posters in
#ICML2022
workshops with amazing collaborators.
3 of them today (22nd) and 2 tomorrow (23rd). See below for details 👇
1
7
26
The University of Copenhagen has multiple open positions in Computer Science, especially in Machine Learning and
#TrustworthyML
Come join us!
Do you want to shape the future of computer science? We are looking for professors at all levels within all areas of Computer Science to join our department! 🚀
📆 Application deadline: 1 September 2024
Take a look at the position and apply here 👉
0
18
36
0
12
24
I have always thought this paper made some of the most pertinent and important points in the space of applied DP learning algorithms. Glad that it finally received the recognition it deserves.
"Position: Considerations for Differentially Private Learning with Large-Scale Public Pretraining," with
@florian_tramer
& Nicholas Carlini got an
#ICML2024
best paper award!
🧵: the personal side of this research, emotional high & lows, & more 👇 1/n
20
30
231
1
1
24
Excited to give a talk today
@EPFL_en
on the impact of inadequate data on trustworthiness of ML algorithms.
I'll talk about some ways how noisy or long tailed data hurts robustness, privacy, and fairness. Much thanks to
@pafrossard
@gortizji
for hosting.
1
1
23
I will be in Kigali where I will be presenting two posters in the main conference track 👇
1️⃣ Theoretically understanding the impact of interpolating label noise in training data on adversarial robustness
2️⃣ investigating the robustness of pre-trained representation robustness
1
1
23
Conferences are so much better when you aren't jet lagged.
1
0
23
After having a great time in Helsinki at the
@ELLISforEurope
workshop on Robust ML (
@ELLISRobustML
), I am now excited to go to Boston for the
#TPDP2023
workshop in Differential Privacy.
If you're around DM me! (Especially if you want DP workshops in Europe too)
In coming weeks, I'll be ✈️ a bit. I'll present posters and give talks on our recent works on robustness/privacy/fairness in ML. If you're interested I'd love to chat (email/DM me)
Dates below 👇
21 - 25 Sep Helsinki
26 - 29 Sep Boston
30 Sep - 3 Oct NYC
4 - 5 Oct Ann Arbor
0
0
34
0
0
21
Call for papers is out for
@satml_conf
. Send in your best submissions once again!
Mark you calendars! We’re thrilled to announce the Call for Papers for the 3rd IEEE Conference on Secure and Trustworthy Machine Learning (
@satml_conf
).
We seek papers on secure, private, and fair learning algorithms and systems
👉
⏰ Deadline: Sep 18
0
20
37
1
4
21
🚀Updated paper🚀
Q: How badly does interpolating label noise hurt adversarial robustness? 🤔
A: We prove a sharp model-agnostic lower bound on adversarial error 😨
and investigate the impact of
✅Inductive bias
✅Noise model
Joint work with
@dpaleka
1
3
20
There should be a Google scholar for number of acknowledgements one receives.
1
0
20
SACs and ACs of
#NEURIPS2023
, please help us out with some statistics of scores in your batch.
Average scores in my review pile looks harsh 😅
1
0
20
2 concurrent papers, one by
@inhaleopenair
,
@minilek
, Lyu, Stemmer and Sarlós () & the other by us solve our open problem presented in
#COLT2022
. (See
@GolowichNoah
& Livni's upper bound)
Interestingly, both have the same hard instance in the lower bound.
Excited about this new preprint with Daniil and
@Kristo
We show a Ω(log T) lower bound for Differentially Private Online learning, even for finite Littlestone classes. This shows a separation between DP and non-DP online learning in mistake bound model.
1
7
47
0
2
20
Come chat with me tomorrow about how interpolating noisy datasets gives a false sense of security (2:30 pm ET on )
If not, we can also chat about Hemingway! On that note I am looking for postdoc in the space of fairness, privacy, and robustness!
Neural Networks can "benignly overfit" some label noise. But is it truly benign ?
Turns out it provably hurts Adversarial Robustness
📢📢Paper: How Benign is Benign Overfitting ? 📢📢
New work with
@puneetdokania
, Varun Kanade (
@ilssontfous
) , Phil Torr
@OxfordTVG
Thread 👇🏾
2
15
72
1
2
19
Collective Action🧑🤝🧑 in ML is the study of the control that a coordinated group can have over ML algorithms.
In our
#ICML2024
paper, led by
@DovOmri
and
@FawkesJake
, we study how the choice of ML Algorithms 🤖 affects the success of the Collective Action.
1
2
17
Looking forward to this workshop and the research visit next January in sunny San Diego 🏖️with
@ccanonne_
@radcummings
and
@bkacsmar
on what to do about making
privacy preserving data science more practical.
EnCORE institute is excited to host a workshop on Holistic Private Data Science for Practice from Jan 8th to 10th, 2025. Check the details here:
@ccanonne_
@radcummings
@bkacsmar
@AmartyaSanyal
0
5
11
0
0
19
Led by
@yaxi_hu
, in collaboration with
@bschoelkopf
, we quantify the privacy degradation due to non-private preprocessing.
Main takeaway is that it can be controlled even when preprocessing introduces correlation between ALL points in dataset.
Paper:
Our work with
@AmartyaSanyal
and
@bschoelkopf
on privacy of pre-processing is online.
Exciting conclusion: Non-private data-dependent pre-processing can be integrated into a DP ML pipeline with minimal degradation of privacy.
Paper:
Visualisation 👇
0
1
16
0
2
18
Honest question: What are the pros and cons of staying in the academia (eg. Postdoc, Research Scientist at Research Labs) as opposed to entering a sector like banking/trading ?
P.S. - Yes, I love doing what I have been doing in my PhD. Also, I am just average at it.
2
1
18
Blog Alert on our recent work led by
@paudejorge
with
@puneetdokania
on
Extreme Pruning at Initialization
Meet FORCE (FOResight Connection sEnsitivity) - extreme pruning of ANNs at initialization by up to 99.9% making network training possible on edge devices. Joint work of
@paudejorge
with
@OxfordTVG
#NeuralNetworks
Blog:
Paper:
2
7
39
0
2
17
We will be presenting our poster today at board number
#708
from 11 am - 1pm in
#NeurIPS2022
on avoiding Catastrophic Overfitting.
Come chat to us
@paudejorge
,
@Adel_Bibi
, and myself about (efficient or not) Adversarial Training and Catastrophic overfitting.
1/n Check out our new work “Make Some Noise: Reliable and Efficient Single-Step Adversarial Training” with amazing collaborators
@adel_bibi
@rvolpis
@AmartyaSanyal
Philip Torr Gregory Rogez and
@puneetdokania
Paper:
Code:
2
7
39
1
0
17
Are you submitting to or have submitted a paper to
@icmlconf
@iclr_conf
@UncertaintyInAI
@AssocCompLearn
@IJCAIconf
#FORC2023
?
Is your paper about trustworthy machine learning with inadequate data/computation ?
You can also submit to our workshop
@TMLunLimited
at
#ICLR2023
!
Excited to announce our in-person ICLR 2023 workshop on "Pitfalls of limited data and computation for trustworthy ML".
We want to foster discussion on how practically relevant statistical and computational limitations can harm trustworthiness in ML.
1
12
34
0
4
17
Less than a month left to apply for this position!
DM/email me if you have any questions.
We have at least one PhD opening (with me) in the domain of Differential Privacy/Robustness in ML.
(with many amazing collaborators/co-advisors inside/outside Copenhagen)
Start: Sep, 24
Location: University of Copenhagen
@DIKU_Institut
@AiCentreDK
.
See
0
35
116
0
4
16
Unexpectedly, I find myself wishing that one of the reviewers for our
#Neurips2023
submission was the Area Chair. 🤷♂️
0
0
16
🚀🚀 Our new work on differentially private dataset release📊🔐 proposes an algorithm to release
1⃣ a DP dataset with (nearly) optimal statistical guarantees for sparse Lipschitz functions.
2⃣ a tight DP "certificate" on its utility for added confidence in the released dataset.
1
2
16
The application deadline for joining us in University of Copenhagen is in two weeks! We have multiple open positions in topics related to theoretical as well as applied machine learning.
Feel free to ping if you have questions :)
The application deadline is approaching! 🔜📆 Don’t miss your chance to apply for a professorship at any level in Computer Science and join our wonderful department. Deadline: 1 September👇
0
3
6
1
2
15
Very humbled to get the opportunity to speak here next week along with many other brilliant researchers
@KAUST_News
0
0
15
This is an amazing opportunity for future PhD students. There's an amazing cross-country collaboration to be had for a PhD in every topic in machine learning.
The
#ELLISPhD
Program is open to candidates from around the world! Apply by November 15 if you are interested in doing a
#PhD
with top
#machinelearning
labs in Europe:
#AI
#ML
#PhDProgram
0
26
62
0
0
15
Also excited to be at
@NeurIPSConf
this year and present this paper.
I'm around till Saturday and keen to talk about all things in Trustworthy ML and check out good restaurants in NOLA.
On my way to NeurIPS where I will present our spotlight paper with
@AmartyaSanyal
, Gizem Yüce,
@FannyYangETH
. Come talk to us (yes, all of us will be there!) during the first poster session (Tuesday 10:45am) if you want to hear more about our adaptive lower bound for SSL.
1
1
30
0
0
14
@togelius
What's the probability that the author of the paper you're reviewing, is reading this tweet and nodding their head in agreement!
1
0
14
For months now, I have thought that "Freundliche Grüsse" is a person and seeing it on every email here, I figured its the most common name is Switzerland.
Today I finally used google translate and 🤦♂️.
0
0
14
We will present our work on using Focal Loss for better calibration in deep networks tomorrow at the UDL workshop in
#ICML2020
.
Amazingly led by
@JishnuMukhoti
@vivekakulharia
with
@puneetdokania
Stuart Golodetz and Phil Torr
Full Paper:
Minimizing focal loss minimizes an
entropy-regularized bregman divergence. Empirically we get better calibration than T-scalng, Brier, lbl smooth, MMCE et . Read our ppr
Joint work w/
@JishnuMukhoti
@vivekakulharia
(1st authors) ,
@puneetdokania
@OxfordTVG
0
4
10
0
4
14
Apart from the researcher's aptitude, what are the other factors, if any, that decides the quality and quantity of the researcher's scientific outputs especially in ML, TCS, Stats, OR etc ?
6
0
14
Its 2020 and I can't believe the amount of space that is wasted on writing the full mathematical expression of softmax every single time it is used even though we conveniently ignore the expressions for the more mathematically ambiguous things that are happening around it.
0
0
13
This was a great experience to give a talk at the Rising Stars Symposium at
@AI_KAUST
and meet some amazing people
@harhrayr
@thegautamkamath
@loiannog
@Karttikeya_m
@yalidux
@oguzhanthefatih
among others
Last week
@AI_KAUST
hosted the 2nd iteration of the rising stars symposium 2023 that is geared towards young researchers, who have recently published significant work at leading AI venues. It was a great opportunity for attendees to discuss and exchange exciting AI research ideas
2
17
35
0
0
13
Finished my
#ICLR2023
reviews one whole day in advance by some miracle. I had a good mix of papers and all were incredibly relevant to what I'm currently actively interested in.
I guess good job
@openreviewnet
and thanks to my ACs.
0
0
13
Excellent example of a red flag while choosing a research advisor.
0
0
13
Had a great time visiting
@krikamol
at
@CISPA
along with
@drjilles
,
@Dziedzic_AD
,
@BurkholzRebekka
, and Franziska.
Thanks Krikamol for inviting me and everyone else who took the time to speak to me, will visit again :)
2
0
13
Very excited to know who is managing the
@iclr_conf
social media account this year 😃
So many great AI conferences these days!
(Of course,
#ICLR2024
welcomes your submission if you're ready to submit.)
2
4
47
1
0
12
When I visited Lorenzo I had a very good learning experience and also lots of fun! Don't miss this.
Other than learning theory you'll also get a healthy (?) dose of focaccia, pesto, and espresso. Don't miss this opportunity.
I'll be looking for PhD students to work on machine learning theory-- statistics/optimization/inverse problems. Learning inverse problems, dynamical systems and implicit regularization are the main keywords!
4
57
218
0
1
12
I find it very odd that some reviewers argue that a method, that works very well, is too simple to be novel.
Surely if it weren't novel but yet productive it would have already existed, right ?
0
0
12
🚀🚀🚀 Job Alert🚀🚀🚀
@DIKU_Institut
is hiring in Machine Learning and Data Science roles.
If you think you would like to join an amazing department in a cool city like Copenhagen, feel free to reach out and see below for more details 👇
0
4
12
Very glad that you could make it!
Had an awesome time at the University of Copenhagen (thanks
@AmartyaSanyal
,
@RasmusPagh1
, and others)! Now off to Aarhus for the Aarhus Summer School on Learning Theory :)
3
1
29
1
0
12
With
@yaxi_hu
, we will present our work on Differential Privacy guarantees despite non-private preprocessing today in
#ICML
at 13:30 hrs. Poster #: 2106
🎉 Very happy that this paper also received a Contributed talk in
#TPDP2024
Led by
@yaxi_hu
, in collaboration with
@bschoelkopf
, we quantify the privacy degradation due to non-private preprocessing.
Main takeaway is that it can be controlled even when preprocessing introduces correlation between ALL points in dataset.
Paper:
0
2
18
0
1
11
@egrefen
@ml_perception
@_rockt
Oh wait I got a better one. "These days the bar of deep learning is very high and this paper just doesn't excite me enough"
And this was a NeurIPS AC after the paper scores were 6,7,7
2
0
12
@RandomlyWalking
@timnitGebru
And while we applaud Churchill so much, how about the approximately 3 million Bengalis in India and Bangladesh who died of starvation because Churchill figured he needed to get some food in reserve for his troops should there be a shortage.
1
2
12
Excited to say that I managed to slept early and I missed it.
I’m starting a support group for people who didn’t see the Northern Lights last night.
We are valid. Stop erasing our existence
Any joiners
This is our flag
2K
5K
67K
0
0
12
We are very excited to host this workshop with an amazing organising team and a great lineup of invited speakers who have kindly accepted our invitation!
Submit your work on how limited data or computation hurts the trustworthiness of ML models!
We will have many interesting invited talks from
@fereshte_khani
@thegautamkamath
@PNetrapalli
, Ruth Urner,
@vitalyFM
And finally, our organising team:
@alexandrutifrea
@AmartyaSanyal
@subGaussian
@FannyYangETH
@fraboeni
@jamiemorgenste1
@jainprateek
@sarahookr
@ilssontfous
0
1
3
0
1
11
Looking forward to attending this workshop in a few weeks to talk about all things robustness and ML.
Excited to co-organise the inaugural ELLIS RobustML workshop at Aalto University, Finland on September 23-24, with a fantastic line-up of speakers!
Workshop website:
ELLIS Programme website:
#ELLISRobustMLWorkshop
1
6
27
0
1
10
If you are around, bug
@FraPintoML
with all your questions live TODAY at his talk 1:20 pm and poster sessions 3:40 pm.
1/n 🧵Happy to be presenting our recent work on Semi-Private Learning at
@satml_conf
in
#Toronto
@UofT
with
@yaxi_hu
@AmartyaSanyal
@FannyYangETH
. Work done at
@ETH_AI_Center
@MPI_IS
📜Paper:
🧑💻Code:
1
9
22
0
0
11
🚨🚨 New paper 🚨🚨 "A law of adversarial risk, interpolation, and label noise" with
@DPaleka
Ever wonder how bad memorising label noise hurts your neural networks (or any machine learning model) ?
Come check our work at the
@responsibledec1
in
@icmlconf
#ICML2022
New theory paper out, where we work on the question "How label noise hurts adversarial robustness?"
TLDR: with uniform label noise, classifiers that perfectly fit the training data can't be robust in the large data regime
joint work w/
@AmartyaSanyal
🧵
3
2
6
1
2
11
The
@ukhomeoffice
is simply refusing to refund me the visa fees for a visa app I couldn't make because their online system failed (Even afte ~two month it says "Your payment is being processed") even though the money got deducted from my account.
@guardian
1
4
11
We're excited to present our open problem in online differential privacy in
#COLT2022
(second session).
Yes, there's some tasty treats for solving them! See you in London, soon.
The list of accepted open problems at
#COLT2022
is now online! Two sessions, 7 open problems, and gallons of glory in solving them!
(Yes, glory is a volume)
2
3
33
1
1
11
I still don't get it. Why are there British expats and Indian immigrants ?
0
0
11