Hal Daumé III
@haldaume3
Followers
27K
Following
4
Media
295
Statuses
17K
https://t.co/jIZyWtJ8gS Human-centered AI, NLP & ML. Director @trails_ai. Prof @umdCS, member of @CLIPumd @HCIL_umd, former @MSFTresearch. Fun: 🧗🧑🍳🧘⛷️🏕️.
Joined June 2010
“What would an AI research & education mission that crossed the whole campus look like?”. This is the question posed to me & a working group by.@UofMaryland's Provost Rice summer'23. Now we’re announcing the answer: #AIMaryland, the AI Interdisciplinary Institute at Maryland. >.
#UMD is proud to announce the launch of the AI Interdisciplinary Institute at Maryland (AIM)! This new institute will be dedicated to developing the next generation of artificial intelligence education, technology and leaders: #AIMaryland
3
11
59
How have I lived my life without ever reading this Peter Landin quote?. (guilty as charged.)
10
457
798
.@ylecun replies to Ali Rahimi's "alchemy" and Ali replies to that. v nice!. my opinion is empirical work is great w/o proving things (I do mostly that) but I am quite allergic to the practice of bandage mountains "just cuz" for dealing w flaws, rather than going fixing the flaws
12
264
587
We are very pleased to announce that the #icml2020 Test of Time award goes to. Gaussian Process Optimization in the Bandit Setting: No Regret and Experimental Design. by Niranjan Srinivas, Andreas Krause, Sham Kakade and Matthias Seeger. >
3
86
452
Please join me in expressing admiration and thanks to the #iclr2020 orgs for their amazing pivot to a virtual conference. As a fellow 2020 chair, this was a monumental feat in minimal time. Thank you @srush_nlp @shakir_za @dawnsongtweets @kchonyc @white_martha and all other orgs!.
4
45
436
📢 I'm thrilled to announce our new @NSF Institute on Trustworthy AI in Law & Society!. @trails_ai's premise is: Participation Builds (Appropriate) Trust. 🌐 Details: 💼 Postdoc, Director, . : 🗓️ Events: >
17
65
359
Super happy to announce new major edit of new chapters on bias, str. pred. & imitation; other edits, fixes & figs!
5
125
336
I'm really loving this article on "Rules of Machine Learning" by Martin Zinkevich . some favorites:
2
94
318
Happy ICML deadline day everyone!.May all your error rates decrease,.All your equations typeset prettily,.All your latex errors cease,.All your prose tend toward brevity,.All your code toward release. And hoping for no crashes on CMT!.
6
33
311
aww my advisees are so thoughtful+awesome!!!! as a congrats for promotion they got me the ultimate prof fashion accessory: a scarf w titles/abs of my papers! now i will never forget to self-cite ;). thank you SO much @raosudha89 @amr_n_sharaf @khanhxuannguyen Kiante & @ihsgnef!!
9
12
288
#nips2017 was my 1st nips in 3 years. still a dumpster fire. seriously, get your shit together. there were awesome scientific events, but my joy is dulled by sexualized events and speeches at sponsor parties & sexual assault "jokes" at closing party (and prob more i didn't see).
18
58
270
I never had the "show up to an exam unprepared" nightmare, but almost every year as faculty I have "show up to 1st day of class with no idea what I'm teaching" dream. Not this year, though! I'm excited for the first day of Human-AI Interaction in 59min! .
4
34
233
Glad to hear that OpenAI has rehired Sam Altman, who was not even candid with them, and rid the board of those pesky women. Seems like they'll be well poised to really combat today and tomorrow's challenges of negative impacts of AI on society.
13
25
218
Some of us generally think of stereotypes and other biases as undesirable in NLP systems. We also generally think of common sense reasoning as a good thing. But is there a difference between these two things?. 🤔🤔🤔. (inspired by talks at wikipedia day per recent tweet).
19
39
230
adding to my math for ML teaching list!
This website is amazing, linear algebra with interactive examples. Vectors, matrix, dot product, etc, cool resource for learning
1
40
222
protip for finding places that might need last minute cleanup in @NipsConference papers:. egrep -iw 'clearly|novel|first|best|note that|discover' *.tex.
9
30
216
I'm thrilled that EMNLP 2023 has a track for Human-Centered NLP. This is how I've been self-identifying of late, which has led to a number of confusions and conversations. I'm not an authority, but thought it might be useful to share this experience. I'd love your thoughts! 1/7.
7
34
221
ooooooh check out the spiffy new vowpal wabbit website:. 🐧 .🐇💨. "Vowpal Wabbit is a fast and flexible learning system that empowers you to test and explore real data and find meaningful outcomes to real problems."
4
48
208
Are you sick of #nlproc folks saying NLP is "solved?" Think your linguistics can poke holes in systems? Prove it!
6
139
197
agree, said sth similar last week:. WHY are we doing openai job for them, and PAYING them for the privilege?. esp, why are we bending over backwards to do so when very basic things (# params, training data, etc.) are hidden from us, basically making half of this job impossible?.
A puzzling dynamic I still don't get is impoverished academics/students paying OpenAI to “understand” or “study properties/capabilities” of GPTs while OpenAI enriches understanding of their product for free. Not to mention all the free publicity/revenue that comes with the.
5
24
192
hah, Sergey Levine, after teaching Reinforce, ended w sth like "if u try this algorithm as is, it probably won't work on anything, even grid world." oh how i wish someone'd told me that ~2010 when we wasted so much time cuz this was not obvious from literature :( #dlrl2018.
1
26
200
Is "AGI" a thing, as far as faculty are concerned?. Roughly - no. 40% don't mention AGI at all to students. Of those who do: >1/2 mention it as a marketing term. ~1/3 as a long term goal some researchers have. And only 1/8 (7.5% overall) present it as an object of study. >.
9
22
182
📢 I'm recruiting a postdoc at UMD.🗓️ Starting in 2024.🖥️🙋 Topics related to human-AI interaction.🔜 Apply by May 31, but earlier if possible!.🤑 $75-$80k+benefits, unencumbered (ie not "big project") funding. Link to full post and form for applications in thread. >.
2
69
167
John Langford on "A Real World Reinforcement Learning Research Program" -- basically laying out alternatives to the "games first, real problems later" approach to reinforcement learning research. And also a note about hiring. 😁😁😁.
3
61
184
With a mixture of all the feelings: yesterday was my last day at @MSFTResearch NYC. The thing I love most abt MSR is the people: not just because they're excellent researchers (they are) but also because they welcome ppl who don't know anything about at topic but are excited. >.
4
3
181
wow, I guess I'm late seeing this, but so impressed at the courage of @KLdivergence and @kat_heller for taking a stand!!!.
2
31
174
The AI Interdisciplinary Institute at @UofMaryland (AIM) is hiring. 40 new faculty members. in all areas of AI, particularly:.- accessibility,.- sustainability,.- social justice, and.- learning;. building on computational, humanistic, or social scientific approaches to AI. >.
3
46
176
Hey grad students!. Check out some internship options at Microsoft Research NYC summer 2020:. Fairness++: Machine learning: Computational social science: Economics+computation:
4
62
172
I wish we could stop worrying about whether something is intelligent or whether something is general, and just focus on whether something is useful, whether something helps people, and whether something moves society in a positive direction. >.
4
25
165
perhaps twitter is considering #280character's because they know it'll basically break all NLP tools trained on twitter data :/.
7
51
154
To all who just finished your ICLR submission, best of luck & I hope you get a chance to relax today/tonight/tomorrow--putting together a paper is HARD: you deserve it!. To those whose work didn't quite make it this time, keep at it, pace yourself & best of luck on your research!.
1
5
163
It's no question that researchers making their code available has been a giant boon for NLP. One concern I have is that it seems to narrow what people will try, possibly based on what code exists, leading to far less exploration of the design space. 1/2.
7
15
153
With great hesitation, but after some bad examples this year (& some nudging), I unified my thoughts and many other sets of suggestions for concrete steps workshop/conf organizers can take to ensure a diverse slate of speakers. Additional ideas are 💯!
3
65
147
and right before the paper deadline i tried a bag of ngrams baseline #shortacademichorrorstory.
4
40
152
just learned there's a formal term for something i tell students a lot. i always say that humans are precision machines not recall machines. (we're able to identify, but brainstorming is hard.) this is apparently called "recognition over recall"! yay!.
7
19
149
This is . a good thing??. The consolidation of AI-related stuff into single one-size-fits-all is IMO one of the worst things to happen (at least to NLP) since I started. Being made durable in infrastructure? We're even more hemmed in. Research is for #ExplorationNotExploitation.
This consolidation in architecture will in turn focus and concentrate software, hardware, and infrastructure, further speeding up progress across AI. Maybe this should have been a blog post. Anyway, exciting times.
5
15
138
I'm back on twitter after 14 months off. What a wonderous 14 months :) :) :). Hi all!. Did I miss anything?.
9
2
145
@mmitchell_ai and i'm sure when the men were hiring other men they got told the same thing, right? riiiiggghhhttt????.
1
4
124
As John said, we've decided that #icml2020 will be virtual; while this introduces challenges, we're enthusiastic about possibilities too. Please read our letter to the community: We wish everyone the best in this challenging time.- Dave, Hal, Aarti, John.
ICML 2020 ( will be a virtual conference. We've been having discussions with ICLR folks about their plans, and will be learning from their experience (they are up first). We hope to enable as much of the normal ICML experience as possible, virtually.
2
41
136
this is exactly why---when famous male authors from top institutions argue that speeding up science with preprints is worth the cost of double blind---i keep my hand on my wallet
OMG reviewers with author information were more likely to recommend acceptance of .- papers from famous authors (+76%).- papers from top institutions (+67%).- papers with male authors (+19%). Horrible biases, same papers evaluated differently (probably guilty myself)
2
42
135
Nice "debates" paper by @zacharylipton and Jacob Steinhardt on "Troubling Trends in Machine Learning Scholarship". their four below. If I'm allowed to add my own I'd add failure to precisely define the problem being solved. Nice read, uphill battle :).
Well here goes. Our ICML Debates paper is live: "Troubling Trends in Machine Learning Scholarship". If anyone needs me, I'll be in witness protection. 🙄
3
45
129
me thinking about writing some proposals.
0
6
128
In NYC area and interested in reinforcement learning?. Come to RL Day at @MSFTResearch on Oct 3!. Free to register!. Awesome speaker lineup: Joelle Pineau, @smdvln, Christopher Amato, Mengdi Wang, Philip Thomas, @real_asli, Geoff Gordon and @FinaleDoshi!.
4
55
130
uhhhhhhhh this image.
Why Logistic Regression should be the last thing you learn when becoming a Data Scientist.
10
31
129
Appears to apply even if your university is in "hybrid" mode, if _you_ in particular happen to only be taking online courses. Hopefully directed research counts as in person, but it will definitely take some university lawyers to try to figure this out.
5
18
124
I cannot sufficiently strongly recommend heading to @black_in_ai today, even if just for part of the day. I went last year and it was technically fascinating and mind opening. Thank you so much to the organizers for investing so much time and energy into this!!!!.
2
27
121
One of the best mental-health-while-working innovations I've seen this year is the "no meetings week" we had at MSR-NYC: A directive to cancel all meetings for a that week. Having everyone do it makes it easier, and even though it's still work time, these have been such a relief!.
4
12
119
Great news Tuesday! The always fantastic @red_abebe named one of MIT Tech Review’s 35 Innovators under 35 for her *pioneering* work on “[sheding] light on the unmet health information needs of people in Africa.”. Congrats and well deserved!!!!.
4
20
114
i think we have a problem :/.
#NIPS2018 The main conference sold out in 11 minutes 38 seconds.
6
15
120
it’s a meritocracy.
Beliefs in academia that need to die. Go!.
3
10
114
Workshops are a great place for students (esp those w/o past papers) to submit, learn in a lower-stakes setting, & gain confidence. But if wksp papers are rev'd alongside conf paps, a totally fine wksp paper now gets harsh scores & reviews, demoralizing & hurting those students.
5
3
115
new paper by Margaret Hu, "Algorithmic Jim Crow" in Fordham Law Review (; h/t David Robinson
0
42
108
@geomblog @d19fe8 @jennwvaughan @hannawallach here's the full version: "Improving fairness in machine learning systems: What do industry practitioners need?" conditionally accepted to #CHI2019. feedback very welcome!!!
1
38
115
precisely why things like “not SOTA -> reject” or “SOTA -> accept” are pointless. i always tell students: it’s great to get good results but if the only thing interesting about some work is that it’s SOTA, there’s nothing interesting because soon something else will be.
Quote of the day: "It was SOTA for about a day or so". 🤓.
0
18
111
I will not attend again until the NIPS board ( puts in place a serious anti-harassment policy and enforces it.
7
20
107
Congrats to all on your @NipsConference #neurips2018 submissions!!! 🎉🎖️. Good luck to program chair @hannawallach and co-chairs @hugo_larochelle, Kristen Grauman, and Nicolò Cesa-Bianchi who now get to deal with many thousands of submissions and reviewers. Be kind to them!!!.
5
10
112
at a really nice invited talk at #nips2017; speaker commented at the beginning that he wasn’t the one originally invited, but that the researcher asked first could not attend because she was at home with her baby. we need to make this easier, esp given social norms re parenting.
8
21
103
I'm really thrilled to announce that on September 24, @MSFTResearch is running:. Reinforcement Learning Day 2018. hosted at the Kimmel Center at NYU. We've got a packed day with nine amazing speakers from inside & outside of MSR!. Info & registration: >
8
32
111
this is a nice primer on some standard techniques for estimating causality (maybe useful for those who are, like i was, intimidated by this stuff).
3
23
110
Are you part of a team working on AI or ML products?. Please help us design tools to help create fairer AI/ML systems!. We're conducting an anonymous survey (~20 min) to better understand teams' practices, challenges, and needs around fairness & bias:. >.
9
79
104
Sorry, no, I won't attend your Dagstuhl seminar with an all-male organizing committee. Can spend my energy elsewhere
4
46
100
🚨🚨🚨. The FATE group at @MSFTResearch NYC is hiring a postdoc whose work focuses on the societal implications of AI&ML!. 📢📢📢. Does that sound like you? Come join us!. Group: Job Ad: Happy to answer any questions you may have!.
2
32
105
Convince me I'm wrong: Generative AI is the new name for structured prediction. An interviewer asked for a def of GenAI & offhand: "an AI system that generates a complex output at once (vs a single prediction)". I later realized that's ≈identical to the def of SP I'd give ~2005.
20
7
105
Today is the official first day of @trails_ai: the Institute for Trustworthy AI in Law & Society!. 🎉🥂🥳. It was a long path getting here, and I'm really excited for the next 5+ years working with an amazing team of scholars, educators and communities. >
3
18
104
Today, I joined 50+ technical experts to call on DHS to drop its “Extreme Vetting Initiative” - a plan to develop (flawed) machine learning systems to flag people for deportation and visa denial based on broad, undefined criteria. #DigitalMuslimBan
1
33
96
New postdoc opportunity in #hcai, people + AI, AI/ethics, and related topics!. We've got a new center at UMD on Values-Centered AI, and want *YOU* to apply for our postdoc position. Deadline: Mar 31, 2023.Job ad: App form: >
3
54
100
I'm thrilled that ARR now finally has a Human-Centered NLP track!. Now we just need to update the contribution types to recognize the kinds of work that go in to HC-NLP, like design, critical computing, understanding people, and computational interaction (taken from CHI).
2
13
103
Some thoughts for ML/NLP readers/reviewers when considering work that does qualitative studies of tools/systems rather than the "my numbers are higher than previous numbers" results that we are used to seeing in our papers. 1/9.
1
33
101
US prof colleagues: putting together class schedules is neigh, and when doing so, I'd encourage you to think about Tue 6 Nov, election day. What you decide: cancelling, normal class, or exam, will impact some students' ability to vote. >.
5
35
95
Hearty congratulations to @Triiiiiista who graduated yesterday after defending her dissertation "Towards Effective and Inclusive AI: Aligning AI Systems with User Needs and Stakeholder Values Across Diverse Contexts" back in March. She starts a postdoc at UT Austin in the Fall!
6
5
97
What, would you say, have been the the most important core innovations in deep learning in the last five years?.
34
32
98
Roses are \TODO{cite someone??}.Violets are \TODO{what's our result?}.Sugar is sweet.And so are you \TODO{consis w data?}.#academicvalentine.
1
20
89
today's linguistic oddity:. We have the same shirt! = not actually the same shirt. We have the same father! = actually the same father. We TOTALLY have the same father! = not actually the same father.
2
4
89
Which of the following words do you think is most rare in US English?. Can you give me a *rarer word* than these (try it, then look it up on google ngram viewer). And what does this have to do with large language models like ChatGPT, BARD, etc.?. Read on. >.
22
21
94
as we come into the final stretch for NeurIPS & EMNLP subs, i recall reviewer feedback got once:. "Other approaches don't have to be bad in order for your approach to be good.". i'd go further: if you're building on other work, it's *better* if it's good: shoulder of giants, etc.
1
18
94
started writing C++ again after a long python hiatus. noticed:.1) it's actually not that bad :P.2) my fingers have lost the ability to automatically type ";".
5
2
92
This is an amazingly clear tutorial on do-calculus for anyone who has (like me) been intimidated!. (also, "all we do in ML is curve fitting"--sure, but a bit reductive/holier-than-thou; the stuff that fits the curve mapping 日本語 speech to English text is not exactly nothing :))
Judea Pearl claims all we do in ML is curve fitting. I wrote this post to explain that claim and introduce the basics of causal inference to ML folks. Machine Learning beyond Curve Fitting: An Intro to Causal Inference and do-Calculus.
1
16
91
how's that arxiv speeds up science thing going?
Ok, so it seems like we’ve established that all we need are vanilla GANs and vanilla LSTMs with well-tuned hyperparameters. All that’s left are getting rid of all those CNN hacks now?.
6
14
92
🤙 Calling #HCAI researchers!.📃 Please submit to the #HCAI @ #NeurIPS2022 workshop!.🗓️ Deadline 09/22/2022.➕ Would love to see submissions across topics like theory, experiences, design, fairness, privacy, transparency, accountability, governance, etc.
1
33
93
I think conference orgs need to set policy for LLMs in experiments. Is NOT using proprietary systems (GPT, Claude, . ) a major concern?. Is ONLY using proprietary systems a major concern?. In reviews of my papers, and as AC, I've seen both, and it's super unfair to authors.
3
7
93
congrats to the latest ACL Fellows!!!. Regina Barzilay. Ralph Grishman. Lillian Lee. Diane Litman. Fernando Pereira. Stuart Shieber.all well deserved!.
1
6
88
good time to point to our maçarico toolkit for implementing learning to search algorithms in #pytorch on top of several structured prediction or RL tasks. go have fun!. much work due to @xtimv, @amr_n_sharaf and @xkianteb. (see
@bgalbraith thanks for the nudge! we were letting the perfect (or "pretty good") be the enemy of the satisficing :(. the reslope code is now public. you'll need two repos:. base implementation experiments ht @amr_n_sharaf, @xkianteb & @xtimv.
2
20
91
i’d rather talk about the best reviewer advice(s) i’ve ever gotten:. 1. others work doesn’t need to be bad in order for yours to be good. 2. make sure at least one baseline has access to the same information as your model to disentangle model from data
What's the worst piece of peer review you’ve ever received? Six academics share theirs.
2
14
89
cs faculty: lol y’all in industry doing coding tests in interviews don’t you know that’s not representative of the job and also super biased?. also cs faculty: i will proceed to evaluate this candidate based on how well they can think on their feet about random research questions.
1
8
87
The opening line alone very clearly presupposes that the goal of AI is to _supplant_ rather than _support_ humanity. This is NOT an inevitability. This is a CHOICE. Already in sentence 1, this doc has failed to consider the most basic question around AI safety. >
3
11
80
If someone has already done this, let me know and I'm happy to join in. If not, please consider signing this letter to the *ACL boards. Will be sent end-of-day tomorrow. #naacl2018.
9
42
83
because it’s not totally obvious from the quoted part, this is about the totally bizarre ecosystem we have created in which, for example, phd students have their names on 50 papers by the time the graduate, which to me is just bonkers and a negative signal.
Since we just wrapped up an AI megaconference, it felt like a good day to plead for fewer papers.
4
9
85
Hey we have this cool paper published that you've probably never heard about before :) :) :).
"Datasheets for Datasets," by @TimnitGebru, @JamieMmt, @BrianaVecchione, @JennWVaughan, @HannaWallach, @HalDaume3, and @KateCrawford, proposes that every dataset be accompanied with a datasheet that documents its provenance, uses, etc.
5
13
80
me selecting reviewer areas (left) when i first started reviewing versus (right) today
4
0
85
wait really?. NAACL 2018: New Orleans.NAACL 2019: Minneapolis.ACL 2020: Seattle. i was surprised that '19 will be in the US, but now '20 too? i'm disappointed that discussions around "maybe alternate out of US" to help with visa issues didn't go anywhere. bad call, imo.
4
15
80
aaaaaand WHO is the absent minded researcher who didn’t actually register?. couldn’t possibly be me, right?.
just landed in Glasgow for #CHI2019 — my first CHI ever!!!! looking forward to seeing some folks I know and meeting *lots* of new people! Now I just need to figure out whether to go to the HCI+ML workshop or the HCI+AI workshop today during the time I’m able to stay awake :)
6
2
80
is it too late to rename probably approximately correct to usually somewhat wrong?.
4
1
80
I've seen this fly by on Facebook a few times and love it every time I see it :). Now it's printed at my desk! Hopefully you like it too!
2
13
79
Exactly this, and it's hard not to think that this is the point. Short thread below; tldr: If you got a faculty(like) offer in CS in the past 10+ years, it's 100% because you're qualified for it & deserve it *especially* if you're a woman, minority, or both. No question. >.
Pedro’s comments are so damaging. I’m applying for faculty jobs now and it feels like even if I get a good position, whether I *deserve* to hold that position will always be in question. I will continually have to re-prove that I am as qualified as my male colleagues.
2
8
81