Daniel Lowd
@dlowd
Followers
3,017
Following
607
Media
208
Statuses
8,690
Machine learning; probabilistic, adversarial, and relational models. Spouse of scifi/furry writer @Ryffnah . he/him
Eugene, OR
Joined April 2009
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Oprah
• 344449 Tweets
Gus Walz
• 282579 Tweets
El TSJ
• 259534 Tweets
Marçal
• 154757 Tweets
sabrina
• 103650 Tweets
Golpe de Estado
• 85683 Tweets
Beyonce
• 77957 Tweets
Barron
• 74265 Tweets
#النصر_الرايد
• 65751 Tweets
Happy Anniversary
• 63418 Tweets
A'TIN ANG PANALO
• 54066 Tweets
Ann Coulter
• 48414 Tweets
Eagles
• 44947 Tweets
رونالدو
• 44484 Tweets
Lugano
• 36073 Tweets
Dotson
• 25721 Tweets
Servette
• 25449 Tweets
Mike Pence
• 20738 Tweets
Mudryk
• 20080 Tweets
#LUGvBJK
• 17641 Tweets
Jermaine Jenas
• 15749 Tweets
Dinesh
• 14884 Tweets
Guiu
• 11268 Tweets
Xmail
• 10920 Tweets
Semih
• 10398 Tweets
Pinned Tweet
Dear Reviewers:
Don't reject because we made it look easy.
Making it look easy is hard work.
Don't reject because our solution is simple.
Simple solutions often have the most impact.
Don't reject because you can imagine additional experiments.
There's a page limit.
30
794
5K
I just had to email my 16-year-old's teacher to explain that he did not use AI for an assignment. (I watched him complete it!)
I also included multiple references for why no one should be using AI detectors in education.
TurnItIn is making $$$$ by lying to schools.
82
312
3K
Machine learning (ML) is a subfield of AI dedicated to the study of MNIST, CIFAR, and ImageNet.
30
112
1K
No, you don’t have to work ridiculous hours to do meaningful work in AI, and it’s harmful to normalize the idea that long hours are necessary.
I’m really disappointed to see OpenAI endorsing such exclusionary practices.
12
102
879
I would like to publicly disavow and distance myself from these comments by my PhD advisor and collaborator.
I have worked with Pedro on a number of projects, and I respect his insight in some areas, but his rhetoric here is both false and harmful.
20
41
772
TFW your former PhD advisor compares you to Stalin and then complains about internet bullies.
15
12
543
Can you remove training data from a random forest?
...without retraining from scratch?
…and still obtain exactly the same model?
…and do it efficiently?
Our upcoming ICML 2021 paper shows that the answer is YES! 🧵
13
36
432
Dolphins achieve super-human swimming ability in < 24 hours! It's just a matter of time before they surpass humans in every task.
#DolphinZero
5
119
385
I'd like to introduce CP-GAN, an amazing new method for generating realistic images!
CP-GAN source code:
cp -r training-data/ GAN-output/
6
89
343
Update: After watching Pedro dig his hole deeper and deeper, I no longer respect his insight in any area.
I would like to publicly disavow and distance myself from these comments by my PhD advisor and collaborator.
I have worked with Pedro on a number of projects, and I respect his insight in some areas, but his rhetoric here is both false and harmful.
20
41
772
7
17
323
Deep learning revolutionizes everything! For example, before deep learning we had double-blind peer review. Now we have arXiv + tweetstorms.
5
91
269
Is a rock conscious?
No.
Is a boulder conscious?
Of course not.
But a really, really large rock, like an asteroid or something?
Maybe! We just don’t know!
23
10
217
My prior belief was that Bayesian methods for deep learning were not yet very practical or useful.
Then I watched this short (11min) talk.
Based on the evidence, I’ve updated my beliefs, and I’m excited about Bayesian DL!
Check it out if you’re curious about uncertainty in DL.
In our
#NeurIPS2021
paper (), we introduce laplace-torch for effortless Bayesian deep learning. Despite their simplicity, we find that Laplace approximations are surprisingly competitive with more popular approaches.
7
95
435
5
23
174
Power corrupts.
This is why it’s bad to concentrate more and more power in the hands of a few — they will use it to exploit those with less power.
Thus, technology that further concentrates power hurts the marginalized, even if it offers some benefit to all.
4
38
168
Who called it “computer security” and not “defense against the dork arts”?
2
33
163
In an effort to combat the rapidly growing review burden for ML conferences, I submitted ZERO papers to NeurIPS this year.
You’re welcome, PC members!
3
9
161
Based on figures I've seen in machine learning papers, "Our Method" is a very effective technique across a wide range of tasks.
4
12
156
Protip: If your deep neural network won’t converge, try remodulating the shield frequencies and reconfiguring the plasma relay to emit a rapid EM burst directly at the subspace anomaly.
9
25
151
@theswayambhu
This was a small assignment (short answer questions in an online biology class), and the teacher is willing to regrade it after I sent the email. So the actual impact on us is small this time… but it’s a big problem overall.
I’m really sorry that happened to your son.
2
1
149
I’m teaching graduate data science this term. Students asked about the Google fiasco on the first day of class.
Concrete harms of the current approach Google is taking extend well beyond Twitter. You'll find this discussion in every Slack channel: university Slack, global Slacks for faculty candidates. We discuss it one-on-one. Students ask us questions about it privately.
1
3
43
1
20
140
TFW a paper you really believe in is rejected from 4 conferences in a row.
6
0
135
@pmddomingos
@RaiaHadsell
My point is that papers are being rejected based on this “scientific review”, and I object to that. The papers’ scientific merits should be evaluated by the readers, not pre-emptively by what is effectively a censorship board.
2
1
126
@pmddomingos
@AnimaAnandkumar
Anima has made some people afraid to express racist and misogynistic views without consequence. That’s a good thing.
You’re trying to make her afraid to work towards inclusive ML. That’s a bad thing.
15
6
125
In case anyone was confused: I am not a grad student. Pedro was my PhD advisor 2003-2010, and I now have tenure at the University of Oregon.
2
1
116
Interesting result for outlier detection in deep learning:
Outliers often end up in the MIDDLE of the distribution instead of outside it!
(See thread for more.)
@roydanroy
Here is a UMAP visualization of held out data from a ResNet trained on 6 inlier classes from CIFAR-10 (various shades of green) with 4 outlier classes (lightest green). The light green points are mostly in the middle
5
5
47
5
14
113
@mark_riedl
In contrast, people like
@AnimaAnandkumar
and
@timnitGebru
face constant harassment for speaking out, and many junior researchers and grad students can’t safely speak out at all.
4
10
104
@sAliAlavi
I didn’t see it myself, however:
1. We mostly discussed research, not politics;
2. Academia paid (even) less attention to diversity and justice;
3. I didn’t see how he was treating other people; and
4. I had a strong incentive to ignore or rationalize anything I saw.
5
0
103
Google leadership has been working very hard to prove that humans are less ethical than AI.
0
10
101
Dear fellow men:
Abusive men often hide their worst behavior from us, and we’re conditioned to ignore or excuse the behavior we do see.
Therefore, when women (& other non-men) tell you that a man is a problem, LISTEN TO THEM.
4
14
97
Thanks to this Turing award announcement, neural networks can now finally get the attention they deserve!
5
9
93
It’s hard for men to understand the kind of harassment women (and other non-men) face on Twitter. It’s a game changer.
Starting to understand why someone would make the mistake of sharing a block list after a bunch of trolls come after her account. (I'm not going to, just like, who even are you people and why do you care so much what I say 😂)
8
0
44
9
7
94
I unironically love this dismissive phrase from the second edition of “An Introduction to Statistical Learning, Second Edition”:
“…niche problems such as image and video classification, speech and text modeling.”
Vision and language are rich domains, but they’re still niche!
2
11
88
For AI/ML people curious about Mastodon — check out !
It’s fast-growing, focuses on AI, and has a good harassment policy. It’s run by
@gradientpub
, so I have hopes it will be managed well long-term.
You can find me there too:
6
10
83
Note that this harassment campaign is specifically targeted at vulnerable people.
I made my position quite clear (including occasional profanity for emphasis), yet received almost NO harassment!
Now why could that be? 🤔
0
10
76
I’m still looking for a few good PhD students!
I work on adversarial and explainable ML, studying how model predictions can be traced to their training data and how we can use that to guarantee robustness.
Interested? Application deadline is December 15, so there’s still time!
3
29
70
Unpopular opinion: the best thing young people can do early in their careers is to know the rules, and so do I — a full commitment’s what I’m thinking of, you wouldn’t get this from any other guy, I just gotta tell you how I’m feeling, gotta make you understand.
2
8
72
So here’s my offer: If you see junior researchers with non-anonymous accounts being targeted unfairly, contact me.
I can’t promise to make things better or agree, but I’m willing to engage in good faith to arrive at a better understanding.
7
4
71
Proud to introduce “Library of Babel,” a completely UNBIASED dataset for training large language models!
KEY IDEA: instead of being biased towards text from one source, we include all possible text!
To get a copy for yourself, just run:
cat /dev/random > library_of_babel.txt
6
3
69
It’s been a challenging week for many, so please enjoy this trolley non-problem.
0
10
70
The patriarchy is like a trolley that crushes countless women, over and over, but we dare not pull the lever to change anything because there are 2 or 3 men standing on the other track, and they *could* move but they don’t want to.
3
12
65
If you’re looking for work in AI, machine learning, or any kind of tech, here is a company to AVOID.
If you already work for
@ThompsonReuters
, now is the time to speak up for ethics and humanity.
Don’t be like pre-WWII IBM, building counting machines for Nazis.
Looks like a division of Thompson Reuters will win ICE contract to assist in the "arrest, detention and removal of criminal aliens from the United States"
12
230
142
2
37
66
Perfect real-world example of data poisoning!
Performance artist generates virtual traffic jams in Google Maps by pulling a wagon full of smartphones
288
10K
32K
0
16
62
@albertgarcia93
@likeavass
@timnitGebru
@ylecun
Sure. Many times, when people address both sides of a conflict as equal, they ignore the significant differences between the two sides. This can have the effect of reinforcing the status quo or promoting bad faith actors. 1/N
1
9
62
@ArcusCoTangens
@timnitGebru
@adjiboussodieng
Racism is everywhere, in everything. It’s like corn — if you’re not allergic to corn, you won’t realize that corn products are used in table salt and bottled water.
Fortunately, corn allergies are rare.
Unfortunately, victims of racism are common.
2
10
62
@pmddomingos
@AnimaAnandkumar
Stop trying to cancel gay rights.
Stop trying to cancel women’s rights.
Stop trying to cancel democracy.
Until then, you’ll have to live with the unspeakable punishment of some people not liking you.
4
7
61
@pmddomingos
@AnimaAnandkumar
Anima has done a lot to make ML more inclusive. I support her, and I appreciate that she doesn’t put up with harassers or people who support them. I also appreciate that she listens and will change her mind when called for.
7
2
58
Many white men treat “racist” and “sexist” as the most damning curses in the language, rather than as attributes that apply to most of us to some degree.
Am I racist and sexist? Probably! Let me know how, and I’ll try to do better.
Furthermore, pointing out someone else's racism, misogyny, misogynoir, gaslighting is not a personal attack against them when it's TRUE. And there are plenty of receipts here.
2
5
69
4
4
58
I have PhD students from Iran. They're great. Their research will lead to better prob models and understanding of ML vulnerabilities ML. 1/7
3
24
56
The bar for privileged white guys is so very, very low, and yet so many of us fail to meet it so often.
(That includes me, but I’m working on it. We all need to work on it.)
1
0
56
CMU's "Never-Ending Language Learner" AI begins to ponder geopolitics, state-sponsored violence, and the ethics of warfare. 🤔
2
16
53
@newplatonism
@AnimaAnandkumar
Yes, there are consequences to expressing unpopular opinions. However, a Twitter fight is nothing compared to the power that employers, academic advisors, and postdoc supervisors hold over junior researchers. If you oppose your supervisor, it can end your career.
1
2
52
Today in the USA we celebrate independence, which means your beliefs are not influenced by evidence: P(A|B)=P(A)
#probability
#alsopolitics
1
13
54
Our work on adversarial examples for text is now up on arXiv!
Ebrahimi et al., "HotFlip: White-Box Adversarial Examples for NLP"
0
21
52
I’m circulating a letter to the ACM to declare that puppies are nice. Please sign to show your support for puppies and also my crusade against the Deranged Leftists who obviously hate puppies. If you don’t sign it, then you hate puppies, just like all my enemies.
4
7
53
@mark_riedl
Thank you, although to be honest, it didn’t take much courage. His behavior is outrageous, it harms the community, and I face few consequences for denouncing it.
1
1
52
@srvmshr
@kate_saenko_
“I understand they have been wronged” — yes, so let’s speak up! They were wronged! That’s bad!
“...our job is not to fuel a fire/cancel culture.” — and since these brilliant women were FIRED for doing their jobs, we should speak up!
2
1
50
Guess the secret word!
Your only feedback is word2vec similarities.
It’s like playing Wordle against a stochastic parrot.
3
12
47
@pmddomingos
@AnimaAnandkumar
As I recall from the earlier conversation, you see no harm in using ML to label trans people. Well, I have family members who are trans, and I care about them and the harms that will be done to them in the future.
4
3
50
Why does this matter?
If you can “unlearn” examples and remove their influence from your models, then you can fulfill the “right to be forgotten.”
If you can do this quickly, then you can handle a stream of requests.
And RFs are very popular and effective. 2/5
3
1
49
Thanks to spammers, we know how to get around this! Try: ”vu1nerab1e”, “e.n.t.i.t.l.e.m.e.n.t”, “D|\/€R$!T¥”, “trans gen der”, “fœtus”, etc.
#spammingforsocialgood
2
9
48
Apparently, GMail thinks an email from the Provost with subject "Promotion Congratulations" is likely to be spam.
...also, I got tenure. 🙃
6
0
49
@TaliaRinger
Live dangerously — write everything in second person!
“In this section, you propose a more efficient algorithm…”
“You would like to acknowledge funding from…”
You’ll spend more time reading your own thesis than anyone else, so it’s mostly correct!
1
0
49
@akilone
@AnimaAnandkumar
When someone refuses to consider evidence or engage in good faith, their world-view is divorced from reality and their “insights” are untrustworthy.
2
3
46
@albertgarcia93
@likeavass
@timnitGebru
@ylecun
And in doing so, he communicated to thousands and thousands of other ML researchers that it's ok to treat Timnit and her research as something to be defended against, disproven, debunked.
To top it all off, he then accused her of not engaging and implied she was irrational. 10/N
1
1
48
@pmddomingos
@AnimaAnandkumar
So this is personal. This is about my family. This is about making ML a place where people I care about are welcome.
1
1
46
I’m pleased to announce that AlphaDreidel has reached human parity in the classic game of dreidel!
The key to AlphaDreidel’s success is our novel closed-form solution to Nash equilibria in games with only 1 action.
Full details in an upcoming tech report.
3
7
47
@pmddomingos
If there is a “woke mob,” I’m probably part of it, so I guess I’m both Stalin and Stalin’s victims. 🤷♂️
Good luck building your anti-woke mob to cancel Anima. Feel free to cancel me too.
4
0
47
@srvmshr
@kate_saenko_
Your stance is that we shouldn’t hold Google accountable until we know everything. Well, it’s been months. If Google had a side to share, they’ve had time to share it.
1
0
44
You know, if your PR team creates this much bad PR, maybe, just maybe, *they’re* the ones who are bad at their jobs and should be fired?
(But we all know why they won’t be.)
1
0
46
@BecomingDataSci
For the record, I’m ashamed he was my advisor, I’m ashamed I kept publishing with him after my PhD, and I’m ashamed I didn’t realize the full extent of who he was until a year ago.
4
1
45
@pmddomingos
@AnimaAnandkumar
No, you don’t.
Countless people have tried to explain things to you, and you refuse to listen.
You support policies that hurt people. And when we object to your stance, you whine about being canceled, as if people disagreeing with you were the worst punishment.
2
1
45
@pmddomingos
@RaiaHadsell
Wasting readers’ time is one harm that we try to avoid in the scientific review process, but it need not be the only harm that we try to avoid.
2
1
41
If I fit a distribution P(X) using fully observed data {x1,x2,...}, what type of learning is this? Supervised? Unsupervised? Self-supervised? Other?
What if I fit P(X,Y)?
What if I fit P(X|Y)?
13
8
44
I know that some people are worried about collateral damage in
@AnimaAnandkumar
’s fight against bigotry — what if grad students and junior researchers who are making an honest effort are targeted due to a poorly phrased opinion?
10
1
45
@BecomingDataSci
Fortunately, he is no longer employed at UW and hasn’t made any meaningful research contributions in a few years. So he’s not really in the research community.
Unfortunately, his prior credentials give his voice far more weight than it deserves.
2
2
43
I love that this is a real figure from a real machine learning paper with >300 citations:
3
7
41
The Pandemic Parent’s Pomodoro Technique 🍅:
1. Pick a task
2. Set a 25-minute timer
3. One child is suddenly hungry
4. Other child is now crying for some reason
...
12. Dogs are eating something inappropriate
…
38. What were you working on?
…
78. Take a 5 minute break
5
2
44
Also, I’d like to propose a new subfield of computer science:
The Machine Learning
It’s kind of like machine learning, but newer and better, and with a different director.
Beyond Suicide Squad, there are so many movies that could be redone and improved by just adding “The” at the beginning….
“The Batman”
“The Groundhog Day”
“The Citizen Kane”
“The The Wizard of Oz”
Etc.
1
0
9
1
5
42
@pmddomingos
@AnimaAnandkumar
But you somehow believe that I only hold these views because I’m afraid of a Twitter mob, as if I were so weak minded as to have no will of my own.
You have so little respect for my intellect and independence. It’s a wonder you ever signed my dissertation.
2
1
43
@pmddomingos
@RaiaHadsell
For example, if you publish a trans-phobic paper (e.g., detecting binary sex based on computer vision), that sends a message to the trans community that they are not welcome in ML. That’s a big harm to the scientific community.
3
2
44
As he explains in his own article, arguing with Pedro Domingos is exactly what he wants, and never ends well for anyone.
Don’t engage, and please pass this message on to others. (Ideally privately, to avoid harassment and signal-boosting his article.)
0
4
42
Can Facebook use AI to fight online abuse? In this article, I discuss the challenge of vague definitions, evasive adversaries, and data poisoning in the fight against spam, hate speech, and more.
2
12
41
I was working to compose a statement like this, but Zachary has said it better. I endorse his statement below.
I’ll say this publicly one final time: that block list was an error of judgment. There are plenty of innocent kids on there caught in the crossfire for liking a single errant tweet. Allies, let’s please never use, disseminate, or defend it. Let’s all move on together.
4
9
271
3
2
41
@pmddomingos
@AnimaAnandkumar
Grades are due tomorrow. I should be grading finals. But here I am, defending myself and my people, broken hearted that someone I learned so much from and looked up to for so long doesn’t respect their humanity.
2
0
42
@pmddomingos
Sorry that I misparsed your metaphor. Would you like to clarify who is like Stalin, so I understand who’s victimizing me right now?
1
0
42
@dan_p_simpson
@timnitGebru
Agreed. I’m skeptical of all large corporations, but
@timnitGebru
was proof to me that people can still do excellent work for the good of all from the inside. I don’t know what’s going on at Google, but it sounds tragic.
1
1
39
The trick to making language models really large is to implement them with vacuum tubes instead of transistors.
2
1
41
We live in a society where male is “default” and female is “other,” and the considerations of men are taken more seriously than those of women.
If you’re a guy, what can you do to make society a little less unfair?
A good first step is to listen.
2
6
41