David Alvarez Melis
@elmelis
Followers
1,758
Following
2,062
Media
24
Statuses
749
Asst. Prof. @hseas @KempnerInst || Researcher @MSRNE || ML + NLP || Previously: @MIT_CSAIL NYU @IBMResearch @ITAM_mx
Cambridge, MA
Joined August 2010
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Vance
• 1490330 Tweets
Carvajal
• 310898 Tweets
Vivek
• 141432 Tweets
Ohio
• 140582 Tweets
America's Hitler
• 124782 Tweets
Pence
• 99239 Tweets
Jack Black
• 84707 Tweets
NCAA
• 58642 Tweets
連休明け
• 53484 Tweets
#WWERaw
• 45341 Tweets
Tim Scott
• 39651 Tweets
Peter Thiel
• 38610 Tweets
National Anthem
• 34735 Tweets
#ラヴィット
• 30546 Tweets
Grammy
• 28691 Tweets
Hillbilly Elegy
• 27261 Tweets
Rhea
• 23953 Tweets
Yale
• 22254 Tweets
#Number_i_ナゲットおトク夏
• 17216 Tweets
#HomeRunDerby
• 12347 Tweets
Appalachia
• 11926 Tweets
副大統領候補
• 11026 Tweets
Update: I've successfully defended my PhD thesis. 🍾🥳 Looking forward to catching up on sleep!
21
0
172
With this announcement out, it's time for an overdue career update! I’ll be joining Harvard SEAS in 2023. I’m beyond thrilled to be part of this superb cohort of incoming faculty, and to join such an exciting, forward-looking and interdisciplinary academic unit as
@hseas
. [1/n]
It's been an exciting year for faculty hiring at Harvard CS: We've recruited 7 faculty members who all have Twitter accounts! Looking forward to welcoming
@ShamKakade6
,
@AnuragAnshu4
,
@elmelis
,
@jefrankle
,
@sitanch
,
@YannaiGonch
,
@viegasf
.
2
15
179
29
5
162
Exactly 10 years ago this day, I arrived in the US with my entire life packed into two suitcases for what was supposed to be just a masters’ degree “and then I’ll head back home” 😅
3
0
51
It’s official! 👩🏻🎓Dr Amores it is. Extremely proud of
@jdthamores
for a spectacular PhD defense. A fitting testament to a remarkably interdisciplinary researcher, her thesis seamlessly weaves together science, art and design. She never ceases to amaze me, and did so again today 🎉
3
0
45
Happening today! Join us in room 220-222 for a day packed with Optimal Transport, from theory to applications, w/ talks by Felix Otto, Laetitia Chapel
@RFlamary
@brandonamos
@ArnaudDoucet1
Florentina Bunea, Smita Krishnaswamy, Sinho Chewi
among others!
#FeelTheMonge
#NeurIPS2023
1
16
44
A propos of
#NeurIPS2022
rebuttals, I asked DALL-E to help me visualize Reviewer
#2
reviewing our paper. Did not disappoint.
2
2
38
Come work with us, now *in person* again! We have projects spanning various areas of ML and Stats, and an awesome team of researchers to work with. Apply below 👇
If you're a PhD student interested in interning with me or one of my amazing colleagues at MSR New England this summer, please apply here
5
68
285
1
6
36
Boston’s ducklings responsibly practicing social distancing *and* wearing face (beak?) masks
#ducklingdistancing
#covid19
1
11
33
This has been a few months in the making, so very happy to finally see it go live! 🎆
The blog gives an overview of our paper (w/
@nfusi
) on OT dataset distances (+ fun interactive visualizations).
The context: we're interested in a notion of distance between datasets... (1/n)
Microsoft researchers propose Optimal Transport Dataset Distance, a method for computing similarity b/w labeled datasets—regardless of whether their label sets are directly comparable. Learn how it can predict task transferability & help data augmentation:
5
176
939
1
4
31
🚀
#OTML
is back for this year’s NeurIPS! We'll have a superb set of speakers (TBA soon) spanning theory, stats, and comp aspects of OT. If you're looking for a home for your work on OT or related topics, we'd love to see it (link to CFP below). Deadline: Sep 29th.
#Monge
✨Another edition of the Optimal Transport and Machine Learning workshop is taking place at
#NeurIPS2023
.
🚀Contribute your work on OT and theory, computation, deep learning, or applications until Sept 29 ()! Details on speakers and schedule coming soon.
2
35
221
0
3
28
Some things sound too good to be true. This is *not* one of them. Been here only a few weeks, but can safely say it’s as great as this listing makes it sound - awesome collaborators, exciting environment, freedom in shaping research agenda, i.e., postdoc paradise. Apply!
We are hiring Postdocs in Machine Learning (among other positions) at Microsoft Research New England:
@MSRNE
2
48
107
0
1
28
𝐶𝑎𝑛 𝑦𝑜𝑢 𝑡𝑒𝑎𝑐ℎ 𝑎𝑛 “𝑜𝑙𝑑 𝐿𝐿𝑀” 𝑛𝑒𝑤 𝑡𝑟𝑖𝑐𝑘𝑠? 🐕🦺
You *can* if you ⟨𝕋𝔸𝔾⟩ your commands!
In this work led by superstar
@JunhongShen1
(w/
@nfusi
@ntenenz
Jimmy Hall) we introduce a simple-but-effective method to ‘repurpose’ LLMs to new domains and tasks
Microsoft presents Tag-LLM
Repurposing General-Purpose LLMs for Specialized Domains
paper page:
Large Language Models (LLMs) have demonstrated remarkable proficiency in understanding and generating natural language. However, their capabilities wane in
0
64
288
2
6
26
If I was a bit younger and knew some biology I would do another PhD with Charlotte. Go work with her, she’s the best!
0
0
25
Happening now! Come hear about this long-in-the-making paper. This was my first project involving user studies, and boy, did I learn a lot. Luckily, I had a super-star team of collaborators with extensive experience on user studies to guide me through it (esp.
@harmankkaur
!)
This week at
#HCOMP2021
,
@elmelis
presents our paper "From Human Explanation to Model Interpretability: A Framework Based on Weight of Evidence" (w/
@harmankkaur
,
@haldaume3
,
@hannawallach
).
We wanted to develop explanations for ML predictions that are meaningful to humans. 1/n
1
13
83
0
3
25
Watching this unfold exactly on the 54th anniversary of the Tlatelolco massacre 🇲🇽, it’s hard not to see the parallels: a regime sensing its grip on power slipping, unleashes its wrath on the youth. Dear Iranian friends and colleagues, we stand with you.
If you're a professor or a student in STEM in a Western country, you probably know someone from the Sharif University of Technology in Iran. They might have be your students, peers, or friends.
That university is now under attack by the Islamic Republic.
🧵(1/16)
#MahsaAmini
164
4K
8K
1
5
24
@natashajaques
@maxhkw
100% accurate 😂. Another classic is the “You might not know it, but *that* is also a kernel” paper
0
1
22
@jacobmbuckman
The fact that so many are missing the sarcasm here says a lot about our field 😑
1
0
20
Heartbreaking. The Ayoloco glacier atop de Iztaccíhuatl volcano in Mexico declared extinct. UNAM researchers left a plaque where the glacier used to be, which (roughly translated) reads:
UNAM declara extinto el glaciar Ayoloco.
La UNAM declaró oficialmente extinto el glaciar Ayoloco ubicado en la cima del Iztaccíhuatl – México, al colocar una placa con un mensaje para las generaciones futuras.
317
8K
17K
1
2
19
Happening soon! Super excited about the wonderful workshop that the organizers of
@ml4data
put together.
#NoLearningWithoutData
(1/8) David Alvarez-Melis
@elmelis
, senior researcher
@MSFTResearch
.
“Comparing, Transforming, and Optimizing Datasets with Optimal Transport” at 8:10 - 8:50am PT.
1
0
2
0
3
19
Happy to share that our paper “Gromov-Wasserstein Alignment of Word Embedding Spaces” was accepted to
#emnlp2018
. Arxiv version coming soon. See you in Brussels!
1
0
18
Filed under cool tools I didn't know I needed: comparxiv ()
Want to compare different versions of the same arXiv paper? (say, vN and vM)
Two steps:
1. pip install comparxiv
2. comparxiv [arXiv Id] N M
Output: a git-diff style pdf with the differences!
0
4
17
Come chat with me!
Join us in the Microsoft booth at
#NeurIPS2022
, booth
#202
, for a 30-minute coffee chat and Q&A session with
@elmelis
, Senior Researcher, Microsoft Research New England.
0
1
10
1
0
17
Highly recommended! I had a great time working with
@jennwvaughan
,
@hannawallach
and
@haldaume3
over the summer. Great team|lab|experience. Apply here 👇🏽.
Calling all PhD candidates interested in fairness, accountability, transparency, and ethics of AI!
Come work with me,
@hannawallach
,
@katecrawford
,
@haldaume3
, Miro Dudik, and the rest of the Microsoft Research NYC FATE group this summer!
Apply now:
8
144
272
0
5
17
A very unique aspect of our lab is that you get to interact closely with folks beyond your main field, such as our wonderful colleagues from the SMC (
@marylgray
@nancybaym
@TarletonG
et al), BioML (
@avapamini
@alexijielu
@KevinKaichuang
et al) and Econ (
@immorlica
et al) groups.
0
1
17
Not exactly the GAN and MLE I meant, but that does sound like a fun trip anyway!
0
0
16
@MSFTResearch
👇 this is a great opportunity for PhD students!! Para mi gente latina: hay una convocatoria específica para estudiantes en (algunos) países de Latinoamérica. Si tienen preguntas sobre la beca, o el proceso de aplicación, mis DMs están abiertos. Fecha límite: 30 de Junio.
1
2
15
🚨 New preprint, work led by
@sunnytqin
Most ML dataset distillation methods focus on creating synthetic data "prototypes" that summarize entire classes. But how much do these prototypes really matter? Not as much as you think! What truly matters will (maybe?) surprise you …
📣 In our latest work, we explore the secret behind the success of many data distillation methods. It turns out that soft labels may be all you need to achieve "data-efficient learning"! (0/5)
@elmelis
1
2
7
1
4
15
Come work with us! If you're interested in working on optimal transport for machine learning and/or transfer learning, we have a couple of projects that might interest you. Follow the link below for details and application form.
If you are a PhD student looking who would like to do a summer internship with me on Program Synthesis (teaching computers to code) or with any of our other fantastic ML folks at Microsoft Research New England, please apply here:
4
51
156
1
4
14
This is happening today! There’s still time to register, so grab a (virtual) seat and see you in a bit!
1
1
13
Postscript: in a year where COVID made the academic job market particularly brutal, an announcement like this one feels superfluous. To all those who weren’t as lucky: stay strong, don’t let the inherent randomness of the process bring you down. I’ll be cheering for y’all! [n/n]
2
0
13
@sp_monte_carlo
Not sure if this is what you have in mind, but the Wasserstein-Bures metric essentially does this (see e.g. , Eq (3)). Even if the distribs are not Gaussian, WB can be used to (non-optimally) 'transport' only their first two moments
1
2
13
This was a fun and exciting project led by two stellar students:
@FrederikeLubeck
and
@_bunnech
(who played both first- and senior-author roles here). Charlotte is in the job market for faculty positions. Any university will be lucky to have her!
Excited about neural
#optimaltransport
methods for modeling
#singlecell
perturbation responses but wondering how to integrate cell birth and death?
@frederikelubeck
has the answer for you in the
#AI4science
workshop
@NeurIPSConf
! Join us in rooms 388 - 390.
2
27
185
0
0
13
Honoring the "NA" part of its name, NAACL is finally coming down to Mexico. Looking forward!
#NAACL2021
0
1
13
@sucholutsky
Very nice survey! I was a bit surprised that Optimal Transport is not mentioned, since alignment is its bread and butter! Lots of OT alignment work in vision, NLP, bio, cogsci and across modalities.
2
0
11
This looks super interesting. And as if the topic weren’t enough, it will be held in my dear
#MexicoCity
, so all the more reason to submit/attend. Kudos to the organizers for helping broaden NLP beyond the ‘usual’ languages
#BenderRule
I am really happy to announce that our workshop "First Workshop on NLP for Indigenous Languages of the Americas" got accepted. See you at
#NAACL2021
#MexicoCity
. Thanks to the amazing organizing team:
@kelina1124
@a11byte
@gneubig
@lexicutioner
@ivanvladimir
@arios272
#NLProc
7
86
278
0
2
10
Great set of tips, particularly re: working on more than one project, and hedging high-risk ambitious projects with low-risk ones. I wish something like this was handed out to incoming PhD students.
10 Tips for Research and a PhD
I've been asked in the past to provide advice on doing research. Here are 10 tips that worked well for me and will hopefully also be useful to others.
14
230
852
0
0
10
Great overview of unsupervised cross-lingual NLP! Sadly I couldn't make it to ACL this year, otherwise I wouldn't have missed this tutorial, but these slides are👌.
The slides of our
@ACL2019_Italy
tutorial on "Unsupervised Cross-Lingual Representation Learning" with
@seb_ruder
and Anders Søgaard are available here:
2
27
120
0
0
9
@sarahookr
🙋🏽looking for students here, particularly if they’re interested in data-efficient ML
0
0
9
@haldaume3
Agreed. Maybe 'good' from an eng. perspective, but not from a scientific one — the homogenization of methods in NLP drove many (myself included) away from the field. I hope this is just a (useful) pruning stage re: architectures, after which domain-informed ideas will strike back
2
1
8
Less than 17% of Wikipedia's biographies are about women (). Donna Strickland did not have one as of the day she received the Nobel prize in Physics earlier this month (!!!). Kudos to
@mitlibraries
for hosting this.
Join us on Nov. 2 for a
@Wikipedia
edit-a-thon on Women in Computing & Engineering! We welcome both new and experienced editors and editing in any language! Lunch provided
#womeninSTEM
@MITEECS
@MITGradStudents
@MITstudents
0
5
6
1
3
8
“TO FUTURE GENERATIONS
The Ayoloco glacier existed here and receded until disappearing in 2018.
Over the next decades, Mexican glaciers will inevitably disappear.
This plaque attests that we knew what was happening and what needed to be done.
Only you will know if we did it.”
0
1
8
Very proud of
@jdthamores
for this well-deserved achievement! 💥🎉
#rockstar
Very happy to see that the essence wearables that I developed during my PhD are one of the
#EdisonAwards
winners 🥳🥳🥳🥳🥳
Thanks to all the people that helped throughout the process & awesome advisor
@PattieMaes
that made this research possible! 🙏
14
2
78
1
0
8
Never thought I’d get sore thumbs from constantly refreshing +5 news apps, but here we are ¯\_(ツ)_/¯
0
0
8
Very timely launch of this great project: . Keep track of the SOTA on your favorite task (in CV, NLP, etc). Includes papers from
#NeurIPS2018
. Crowdsourced, curated and community driven. Check it out!
0
1
8
Can you adversarially learn a distribution on Euclidean space from samples in a different space and/or format (eg a relational graph)? Come hear
@_bunnech
answer this question early tomorrow, room 517A.
#NeurIPS2018
#R2L
. (Spoilers: you can!)
0
1
8
More broadly, we see this as a small but promising step towards a Data-Centric ML paradigm that focuses on the datasets that feed our models, and which would be orthogonal and complementary to the current (model-centric) one. Stay tuned for more on this...!
0
0
7
In the meantime I'm super excited to continue at
@MSRNE
, where I’ve found colleagues who besides being amazing researchers (and human beings in general!) have been incredibly supportive throughout my decision process. I will be forever grateful to them. [4/n]
1
0
7
I will build a research group focused on machine learning adaptability, reliability, and interpretability. Among its core missions: to provide a bridge between ML and the Applied Math folks at SEAS and IACS, to collaborate with the broader science community at Harvard, and [2/n]
1
0
7
🚨 Two weeks left to apply fo this!
Quedan dos semanas para enviar solicitudes! 🚨
¿Estás en un doctorado (>=2ndo año) en computación (o áreas afines*) en una universidad de Argentina, Brasil, Chile, Colombia, Costa Rica, México o Puerto Rico? Entonces este 🧵es para ti! 👇
Today, we're thrilled to accept nominations and proposals for the Microsoft Research PhD Fellowship across Asia-Pacific, Canada and the United States, EMEA, and Latin America. Program details vary by region. Review your region's tab for info and key dates:
32
752
2K
2
5
7
Come work with us!
0
1
7
Before getting into the nitty-gritty, some pointers:
Paper:
Talk:
Code:
1
0
6
Ever struggle with compilation errors when uploading to the
@arxiv_org
? Tired of patching dependencies to avoid option clashes? .bbl rejected because of version mismatch? Do yourself a favor and check out this super useful tool by
@d00gs
.
#arxivproblems
0
1
6
This is such a great initiative, especially for international students who might not familiar with the process and expectations of US PhD applications. Kudos to the GAAP team.
Applying to grad school in EE/CS this fall? …need help?
Ask the MIT EECS Graduate Application Assistance Program! GAAP pairs applicants who need help with current PhD students for 1:1 mentoring. We match mentors weekly so it's never too late to sign up!
1
41
123
0
0
6
@ChristophMolnar
This doc (compiled by
@vkrakovna
, I think) has various other examples of funky ML behavior due to misspecification, including many in RL.
0
0
6
Nice and thought-provoking summary of the exciting recent developments in Deep Learning theory by
@prfsanjeevarora
. Haven't gotten around to reading these papers yet, but after reading this blogpost they're definitely back up on top of my to-read list.
"Is optimization the right language to understand the brain?" is a famous controversy in neuroscience. My new blog post asks if optimization is the right language even to understand deep learning? (TL;DR: let's think: trajectories!)
6
200
716
0
0
6
... to tackle problems at the intersection of ML and humans/society alongside others who are leading the way in approaching this issue from various angles, from SEAS’ own
@HarvardHci
, to
@HRCS
and
@harvardmed
. [3/n]
1
0
5
Generative Models + Geometry = 💥. This project was a lot of fun. After we hit a few roadblocks,
@_bunnech
did magic and made it work. Check out the preprint!
0
1
5
Sure, it’s a home from which I can be kicked out at a moment’s notice (I’m still in the same type of precarious student visa as that day I landed in JKF 10 years ago 🤷🏽 💔), but a home nevertheless
1
0
5
@FelixHill84
Hottest take: all of the above, plus, advertising the *rate* of papers you got accepted is even worse and serves no purpose but signaling. This of course does not apply to those tongue-in-cheek tweets touting 0/0 acceptance rates
0
0
5
@_onionesque
By far, the most concise, densest, yet beautiful, textbook on probability theory I’ve encountered. A true gem. Was lucky to learn some of that content directly from the one and only Varadhan himself :)
1
0
4
Ok this has officially won the Academic Twitter today
Wow! A great performance by some amazing Women in Theory (inc.
@immorlica
,
@dahlia_malkhi
,
@mkwoot
, Shafi, Ronitt, Katrina Ligett, and many more), lyrics by Avi Wigderson.
Million times better than the (literally) tone-deaf celebrity "Imagine".
5
42
157
0
0
4
Ok, and so what? We show this distance is a good predictor of model transferability across tasks, so it can be used to guide choice of pretraining dataset and/or augmentation protocol in transfer learning. Check out the full paper for details!
Fin! (n/n)
0
0
4
@Aaroth
And then there's the issue of feature "removal". In coalitions, removing players is perfectly well defined, but in most types of inputs to ML models, just "removing" a subset of the features is either not defined (eg images) or moves away from data manifold, making interp. moot
1
0
4
@KevinKaichuang
First learner appears to have reached a local minimum, judging from that awesome car vector field 😅
1
0
4
@srush_nlp
@taolei15949106
Nice! Hadn’t seen this one. Agreed, always amazed by how Tao blends methodological creativity with practical know-how like very few others do. Must read!
1
0
4
0
0
4
@zzznah
@gabrielpeyre
Slightly more recent than those, we explore the idea that 'moving things together is cheaper' via submodularity here: There's also nice work on anchor wasserstein distances (eg by
@CuturiMarco
and colleagues), which also plays on the idea of grouping
0
0
4
@srush_nlp
Lots of great answers already, my 2cents: if (1) is sample-wise, can easily construct examples where both are true (eg “losing the tails”). Even if (2) is distributional and over a rich enough class of “distinguishers”, sample complexity would make precise estimation intractable
0
0
4
@_onionesque
Just read it last week, and it absolutely piqued my interest too! It’s one of the most appealing notions of transferability I’ve come across.
2
0
4
1. Represent classes as distribs over features ⇨ allows to compare any two classes via OT
2. Step 1 + feature space metric ⇨ metric between feature-label pairs
3. Use (yes, you guessed it!) OT again to get distance b/w *distribs* over feat-label pairs (aka datasets) (5/n)
1
0
3
@haldaume3
Here's an entire masters' thesis devoted to this question for the case of Spanish: (disclaimer: my mother advised it, and I couldn't find the student on twitter to tag them here). Happy to help translating if you're interested.
1
0
3
@XandaSchofield
Taro says “hi Xanda, keep it up!” (or at least that’s how I interpret his ‘woof’ 😀)
1
0
3
@srchvrs
@IgorBrigadir
@complingy
@ivrik
Here: we used a method similar to WMD to compare embeddings across languages, yields a distance between them (see Fig 4). Granted, it's highly dependent on corpus + embedding, and essentially semantic.
2
0
3
Maybe you're not into meditating with closed eyes. How about meditating in outer space? On a tropical beach? Passing through a shapeshifting psychedelic piece of art?
@zindelsegal
@JacobSAday
@TRIPPVR
@MeditationVR
@jdthamores
@spiritualbutnot
@MindWise_CK
3
7
23
0
0
3
In fact, we think that *even more* problems in ML can be tackled with this data-centric perspective, but let’s not get ahead of ourselves!
1
0
3
Do you ever feel like a big part of ML (especially in practice) is 𝗱𝗮𝘁𝗮𝘀𝗲𝘁 𝘁𝗿𝗮𝗻𝘀𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻 in disguise? So do we! From generative modeling to data pre-processing, various tasks can be cast as transforming datasets according to some predefined criteria
1
0
3
Very thorough summary of state of AI in Mexico + recommendations. Glad to see such a broad analysis of impact (ethics, gender issues, jobs at risk). An encouraging step towards a national AI strategy in
#Mexico
. .
@CMinds_co
@UKinMexico
@oxfordinsights
0
6
3
We showcase this method for tasks ranging from dataset ‘shaping’ (e.g., imposing geometric constraints) to model ‘re-purposing’ — transforming dataset B so that a (frozen - *including* the last layer) classifier trained on dataset A achieves high accuracy on B too.
1
0
3
@JustinMSolomon
I recently had one with “emotional stability” too and I just couldn’t believe it. Leaving aside the fact that very few would even be qualified to objectively answer such a question, it seems like basing any decision on it must violate some Equal Opportunity criterion.
1
0
3
Our structured black-box interpretability paper (presented a few hours ago at EMNLP) was featured by MIT news
0
0
2