So excited to make ML models more reliable using causal reasoning. LET'S DO THIS!
Also: *could not* have done this without support from my students,
@UMichCSE
,
@michigan_AI
, and UM eHAIL. I am so grateful
Delighted to congratulate Prof.
@Maggiemakar
for receiving the 🌟NSF CAREER🌟 Award!👏
Title: “From Fragile to Fortified: Harnessing Causal Reasoning for Trustworthy Machine Learning with Unreliable Data”.
Very excited to announce that I am looking for PhD students to work with
@UMichCSE
. If you're interested in causality + machine learning and would like to spend your PhD years NOT losing half your income on rent, I strongly encourage you to apply to our PhD program
Happy
#NeurIPS2021
review day! Reminder that in 2016, 3 NeurIPS reviewers were convinced that my work on identifying asymptomatic carriers of infectious diseases “will not be relevant for a wide audience”. Reviews are always imperfect, don’t let it get to you 🙃
#yougotthis
I'm on the job market! I like to think about making causal inference efficient using ideas from ML, and making ML more robust using ideas from causal inference. Check out my work on and reach out if you're hiring & think I'd be a good fit!
FINALLY transitioned over to pytorch after years of tensorflow. Just based on a couple of days of playing with pytorch: better distributed learning, better documentation, BETTER HANDLING OF SLICING.
So much better. Ouff.
[1/n] 11 yrs after moving to the US, my application for permanent residence on the basis of exceptional ability was approved. To the MANY PEOPLE who suggested the only way to do this is through marriage: women can color outside the limited boundaries of your imagination
@suchenzang
UMich prof here! I've never heard of this, my guess is: it's a scam. I can't find documentation or links to this data.
If it was available, it would have been made available to UMich faculty, which it hasn't. Do you mind DMing me the email of the sender? Can check if legit
I successfully defended my thesis! So thankful for the mentorship and support of so many. I believe I am now qualified to prescribe antibiotics
#imadoctor
[1/n] New arXiv paper alert! ML models can take shortcuts, which makes them perform poorly under distribution shift. We used some ideas from causality to get models to use the right signal. W
@davisblalock
,
@packer_ben
, Yoni Halpern, &
@alexdamour
. 🧵
I feel like some day a student will walk in and be like “oh what are these?” And I’ll be like “long long ago, in a far away land, there were in-person conferences where we stood around posters and wore name tags.”
#covidsucks
[1/n] This tweet, while likely well intentioned is misleading. Student data was not and has never been for sale by U of M.
The content referenced includes papers and speech recordings that had been voluntarily contributed by student volunteers participating, under signed consent
@zacharylipton
I always always suggest that people start with John Shawe Taylor's MLSS lectures. Incredibly intuitive, step by step, holds your hand through every line of math.
Stop👏using 👏 anthropomorphic 👏 language 👏 to 👏 describe 👏models 👏.
It can't see --> it can analyze image input
It can't hear --> it can analyze speech signals
It can't speak --> it can produce audible output.
Stop. Please.
🎉 Huge congratulations to
@Maggiemakar
on receiving the
@Google
Research Scholar Award for research on "Reliability through causal alignment: causally motivated models for pain management."
Here's to many more achievements ahead!👏
Happy to announce that I'm looking for PhD candidates to join my lab
@UMichCSE
this year!
If you're interested in working on causality, robustness or healthcare applications apply to our program!
After all, who wouldn't want to join the biggest house party in America? ⬇️
Extra happy that my
#NSF
proposal officially got approved!
Clinical data often reflects biases in access to & quality of healthcare. I'm excited to build causal & predictive models that ameliorate the effect of these data inequalities
#letsdothis
To everybody who's ever told me that it's a great thing that I have an expressive face: I continue to prove that you are wrong. It's strictly bad to have an expressive face. Below is a compilation of me at poster sessions making faces.
#Neurips2022
pic credit
@shengpu_tang
For some reason, every year leading up to the NeurIPS deadline, my imagination runs wild with all the other careers I could have picked.
This year, in my head (obviously), I'm a DJ. I'm all about that Ibiza life.
Happy NeurIPSing everybody.
3 NeurIPS accepts :)
My favorite part --> they were all led by students who worked really really hard.
So incredibly proud of Jiayun Zheng,
@shengpu_tang
,
@SarahJabbour_
and Jiaxuan Wang
More paper details coming at you soon
#NeurIPS2022
I can't be the only person who spends like a solid 30 minutes admiring her own rebuttals after submitting them...
Like dang Maggie, that's such a valid point, I bet it will convince all of them
#whosgoingtotellher
True story: you can get 2 accepts & 1 weak accept from 3 reviewers & get your paper rejected because a meta-reviewer decides to act unilaterally. I'm disappointed as an author & a reviewer who spends time writing reviews thinking they matter
#Faact
@BedoyaFTC
@FTC
I feel this, all of it! I still have the napkin I was given on the plane 11 years ago when I came to the US, it's one of my post prized possessions :)
I'll be presenting our work on Causally motivated shortcut removal at
#NeurIPS2022
poster session 2 at 4pm in Hall J, poster
#323
. Stop by if you'd like to learn more about the work OR if you're planning to apply for PhD programs. I am looking for motivated students to work with!
[4/n] last thing: Immigration is HARD even for people with privilege like me. I couldn't go home to hug my mom when her parents died, or care for my parents when they got COVID, or dance with my family at my cousin's wedding. So please be kind to immigrants. They carry a lot.
[1/n] Haven't heard anybody talking about this, so I'll swallow my disdain for mass tweets & do it myself: the pandemic has wrung us out for very real reasons. Today I am particularly sad about losing the awesome communal feel that comes with physically going to
#ICML2020
[2/n] I don't know how we are going to manage misinformation going forward. It'll continue getting more difficult. But maybe the answer here is to lead with kindness: before hitting the post/share button maybe we can all stop for a second to think about who gets hurt when we do.
"Being a woman in CS can be challenging because you don’t see a lot of women in CS right now.... But having a woman as an advisor is always a reminder that women do have a place here. At CSE, the Michigan AI Lab has a faculty that is 50% women"
@SarahJabbour_
Interested in grad studies at U-M CSE?
We sat down with PhD candidate Sarah Jabbour to discuss how collaboration and connection are centered in her experience as a graduate student.
[1/n] I'll be at
#NeurIPS2022
Nov 28 - Dec 1. Happy to chat if you're considering applying to
@UMichCSE
/
@michigan_AI
. If you're an international student who was unable to come due to visa issues also please reach out. I'm going to make time during the week of Dec 5 for zoom chats
Very excited to announce that I am looking for PhD students to work with
@UMichCSE
. If you're interested in causality + machine learning and would like to spend your PhD years NOT losing half your income on rent, I strongly encourage you to apply to our PhD program
[3/n] It's also hard to overstate the importance of my privilege: My parents were able to financially support me during my very-expensive-by-egyptian-standards undergrad education. I will work hard to make this opportunity available for international women with less resources
Learn more about building efficient, reliable models for decision-making during today's AI Seminar.
@Maggiemakar
talks about "Machine Learning and Causality" TODAY @ 4pm.
Am I misreading this? There were no representatives from academia in the president's conversation about the dangers of AI? If true, we should all be worried about this. Industry entities represent a very specific, biased group of the actors in the AI field
Can academics have work-life-balance? What's it like to advise PhD students? How much grant writing do you need to do as a professor?
Come. Ask us your questions!
📢Students, join us for an ask-all-your-questions panel with
@UMichCSE
faculty on what it means and what it takes to be a professor!
🗓️ FEB. 22 @ 4:30-6:00 pm ET in Beyster
#1690
With
@radamihalcea
, Kim Diaz, Ron Dreslinksi,
@Maggiemakar
, and
@AlansonSample
.
There is nothing more stupid than convincing yourself to start a major code refactor on a Saturday at 11pm. Why did I do this to myself? Why am I like this?
~6 years ago I was a TA for the first time and I met my very first favorite student*
Today
@RaghuAniruddh
successfully defended his thesis and I could not be more excited for whatever he's going to do next.
They grow up so fast :')
* yes I have favorites. everybody does.
[1/n] ICML paper alert! When an infectious diseases is frequently propagated by asymptomatic carriers, labeled data for that disease are typically biased: they capture symptomatic infections more than asymptomatic carriage.
What if a decision maker only needs bounds on potential outcomes under different interventions rather than the most precise estimates? Does that make the learning task easier? Stop by our poster session at 9am or 8pm EST to learn more!
Poster details
My toxic trait is that if I'm super intrigued by one question in one of the projects I'm working on, I cannot for the life of me focus on anything else till it's resolved...
Today, remember the wise words of Lizzo, the patron saint of those who wait for
#NeurIPS2022
reviews:
"In a minute, Imma need a sentimental man or woman to pump me up".
BREAKING NEWS:
The 2021 Sveriges Riksbank Prize in Economic Sciences in Memory of Alfred Nobel has been awarded with one half to David Card and the other half jointly to Joshua D. Angrist and Guido W. Imbens.
#NobelPrize
It's not that I don't appreciate a robust conversation about the implications of large language models. It's just that i might literally combust into flames if someone says AGI one more time 😱
This was probably the most fun I've had in a while.
Listen to this for a good dose of ML + causality, infectious diseases and *most* importantly: laughter.
Thanks
@EpiEllie
and
@LucyStats
for being fantastic hosts.
Things I learned after finishing my
#PhD
: there are no friendships like the friendships you make at 3am on the night of a deadline over stale pad Thai & the smell of your labmates’ (& your) body odor because you haven’t showered in days. These are the BEST friendships ❤️🥹
[3/n] Today was tough. It sucks being accused of selling out our students' data, when we truly care about our students & their privacy. Not to mention many of us can be making way more $ working for companies that actually do trade our data and trample our privacy
I can't be the only one really feeling the weight of the war and covid and all the things this morning. It is probably ok to hit pause for a second. So here is a happy picture for you. Feel free to share happy things with me--dogs, research, food. Mostly dogs though.
My favorite: Our 2015 (read: pre-pandemic) paper on identifying asymptomatic spreaders of an infectious disease was deemed "too niche, unlikely to have an impact on a wide audience".
(It got into AAAI...)
NeurIPS and others should take the same path as TMLR and significantly downweight the role of estimates of impact. We should be focussing on solid, steady, empirical/theoretical advances. We can give 10% of papers one year later "test of short time" awards if they show impact.
[2/n] it's impossible to overstate my parents' contribution in this journey. They valued education enough to allow their young unmarried daughter to travel alone to the US (gasp 🙄), which is insanely progressive for average Egyptian parents
"When people talk about interpretability, they are usually talking about causality, which I think is important and more people should work on it"-Yann Lecun
#NIPS2017
On one level, these allegations are so disheartening. On another level: it makes sense and not at all shocking. This insane ML rat race of publishing at 2-3 specific conferences encourages perverse incentives to publish incremental work & in this case downright cheat (allegedly)
My advisor said the same: “Do what makes you satisfied. If it leads to climbing up the promotion ladder, it’s a double win. If it doesn’t, your disappointment will be offset by the fact that you pursued your interests.”
[1/n] This tweet, while likely well intentioned is misleading. Student data was not and has never been for sale by U of M.
The content referenced includes papers and speech recordings that had been voluntarily contributed by student volunteers participating, under signed consent
I had been REALLY excited for the causality panel but technical difficulties are making it impossible to hear the panelist. Super bummed out
#NeurIPS2022
Excited to announce that our paper on causally-motivated shortcut removal (see thread below!) was accepted for publication in
#AISTATS
! Code is also available here
[1/n] New arXiv paper alert! ML models can take shortcuts, which makes them perform poorly under distribution shift. We used some ideas from causality to get models to use the right signal. W
@davisblalock
,
@packer_ben
, Yoni Halpern, &
@alexdamour
. 🧵
We are organizing an ICML 2022 workshop on Spurious Correlations, Invariance, and Stability (SCIS)!
We have a fantastic set of speakers!
Submit your work (by May 24) addressing spurious correlations: discovery, methods, evaluations, and beyond!
Wouldn't it be super cool to read a paper with a negative result? "Hello world, we tried this and here is why it doesn't and shouldn't work".
#SaturdayNightPondering
Got a great paper on spurious correlations rejected from
#ICML2022
/
#UAI2022
? Just submitted to
#Neurips2022
& worried about the fact that you're playing roulette with the odds being more like 330 to 1 rather than 37 to 1?
Submit your work to our workshop! Deadline: May 27
Now that we're post-NeurIPS...
Consider submitting work to the ICML 2022 workshop SCIS: Spurious Correlations, Invariance and Stability!
We extended the submission deadline to May 27th, all details here:
That awk moment when you wake up from a dream where computer scientists were all getting infected by some disease that made us throw up matrices. The good news is that it wasn't infectious, we were all ingesting something toxic: num pie
#DontEatNumpy
#CovidDreams
In my intro to ML class I was asked "how can we ever calculate the bias/variance of an estimator? That's impossible"
Took me a LOT of energy to not go "YES IT IS LET ME TELL YOU ABOUT BOUNDING THE GENERALIZATION ERROR & ALL THE WAYS YOU CAN MEASURE FUNCTION COMPLEXITY"
#SLTisCOOL
Me recording talks in 2020: impeccable office, 100 different versions of the talk, repeat recording because of minor errors
Me recording talks now: leftover lunch visible in camera, one recording, keep going despite minor errors or occasional sneezing
#imoverthis
Maybe
@DTE_Energy
will listen if I make this rhyme
We haven’t had power for a very long time
Tomorrow is my birthday, I turn thirty two!
If I don’t wake up in my bedroom I will be blue
Give me back my heat
Make my birthday complete
#annarbor
#dte
I said it before and I'll say it again. The best friendships are the ones forged at 2 am when you're pushing hard for a deadline and haven't seen the inside of the shower in a couple of days.
Congrats to Dr. Suresh :)
@MarzyehGhassemi
presenting on biases and harms of spurious correlation in the context of healthcare setting at the
#SCIS
workshop at
#ICML2022
: e.g., your algorithm can tell if the patient is black even if YOU can’t
I agree with Maggie. This language used among CS people is a shortcut since we know what it's actually doing and NOT DOING. But when we use this language facing the public, it is factually misleading to them. I'll add AI doesn't "think" or "know" or "feel" or "learn" in the…
“Ten years ago I was kicked out and forced to retire.”
Our new medicine laureate Katalin Karikó (
@kkariko
) told us how much it means to be awarded the Nobel Prize after a scientific career that has been full of challenges.
Ten years ago, Karikó was still doing all her…
When the pandemic is over, I'm going to visit all my collaborators' houses and demand to see the rest of their office rooms that are outside of the camera shot
#curiositykills
We've launched Layer Health
@layerhealth
, a new AI startup solving healthcare's information problem, with large language models. I'm CEO, working with an amazing team and backed by $4 million from Google Ventures
@GVteam
@generalcatalyst
&
@inceptionhealth
I have got to say: one of the most valuable yet very under-rated skills that I've learned during my PhD is how to create good visualizations using powerpoint. Props to
@frejohk
for the awesome tips.
Thomas Renz: We have incentivized the murder of patients rather than incentivize treatment...When you get admitted for COVID, they get paid more. When they put you on remdesivir, they get paid more. When you get ventilated, they get paid more. When you die, they get paid more.
#tbt
2018 when I (1) had to spend a solid 3 minutes explaining what asymptomatic carriers are, (2) tried to make my research in infectious diseases relevant to an ML audience by saying that you can use it to model rumor spread in social networks