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
The news and videos are just coming out. I will post a few here.
We see students being trapped in a parking lot at the university, escaping from those who appear to be paramilitary.
(2/16)
#IranProtests
#SharifUniversity
I am teaching a course on Reinforcement Learning!
I cover fundamental concepts and algorithms and often prove why they work.
All the slides, lecture notes, and videos are available here:
The YouTube channel for the course is
This all started about two weeks ago when the "morality" police of the Islamic Republic murdered a young woman named
#MahsaAmini
.
I've written about it before to provide some context.
By now, you can find about the issue on mainstream media too.
(5/16)
A few days ago, the morality policy of the Islamic Republic of Iran murdered a 22 year woman called
#Mahsa_Amini
. You can see in this video how cruely they treat women (not Mahsa).
Trigger Warning: Watching this is upsetting.
Mahsa's death has a Butterfly effect. Iranians have since protested in Iran and all over the world almost everyday. There was a demonstration of 50K in
#Toronto
on October 1st. And it wasn't just in Toronto. 150 other cities around the world joined as well!
(7/16)
#MahsaAmini
Since the murder of
#MahsaAmini
, Iran has permanently changed. Not that it was the first time such a crime has happened. But this time is just different.
(6/16)
This is not the first time that the IR has violently suppressed Iranians. Its foundation is based on suppression of people and their liberty. It has happened too many times for me to report.
(8/16)
Within a university, my memory goes back to July 1999, when they attacked the University of Tehran, killed several students and arrested 1000+.
(9/16)
#SharifUniversity
If you have read the novel 1984 by George Orwell, that should give you an idea of how such a government works.
Sometimes I wonder whether they have actually used that novel as the blueprint of how to govern. Or maybe all authoritarian governments tend to be alike?
(13/16)
Going back to today: Raiding a university will probably not be limited to
#SharifUniversity
. This will happen at other universities too, if not already. They are afraid of anyone who thinks and rebels against their indoctrination.
(12/16)
The 1999 invasion of the university started a new era of resistance, and unfortunately also a consequent depression for a generation of students.
(10/16)
Those students started hopeful of then President Khatami (1997-2005) and the Reformists movement, but got disappointed a few years later when much of their hopes and wishes didn't come to a fruition.
The realization for many was simple: IR is not reformable.
(11/16)
#MahsaAmini
I don't know what will happen to Iran in the near future. The IR might start a bloodbath as a temporarily Band-aid to save themselves.
I am convinced that the top officials will not hesitate to do so. Think of extremist psychopaths with political and military power.
(14/16)
The mid-rank government officials or military might be another story. They may evaluate what they are doing and decide to do otherwise, either due to their awakened conscience or a reevaluation of their own utility.
(15/16)
#IranProtests2022
I plan to recruit one or two graduate students in
#ReinforcementLearning
this year. If you are interested, apply through
@UofTCompSci
@UofT
and mention my name in your application (deadline: Dec 1).
If you are interested in graduate studies on reinforcement learning, consider applying to the Department of Computer Science, University of Toronto and work with me.
As part of the Pan-Canadian AI Strategy, we are announcing an expansion of the Canada CIFAR AI Chairs program, bringing the total number of chairs to 46, from 29 announced last December.
Meet Canada’s AI leaders:
#CIFARAI
My PhD students are interesting in doing research internship next summer, so I thought I should advertise them here, in case you are hiring.
They have a solid foundation in
#ReinforcementLearning
and
#MachineLearning
.
Contact them directly, or DM.
Please RT!
"Without a perfect model, model-based RL is hopeless!"
Our paper at
#ICLR2024
challenges this belief! Even an inaccurate model can help a lot. Don’t throw it away!
Title: Maximum Entropy Model Correction in Reinforcement Learning
Paper: 🧵(1/7)
Why do we hold
#NeurIPS2022
in a closed-bordered country such as the USA? This greatly impacts many researchers with less privileged citizenships.
@NeurIPSConf
I think this is a good time as any to remind people that many students from India (and many other countries) with accepted papers to
@NeurIPSConf
wouldn't be able to present because the US Visa situation is so backed up (the next available appointments are in Mar/April 2024).
@TOAdamVaughan
The blame, foremost, goes to Mr. Vuong for not disclosing his past. That was immoral.
But part of the blame goes to the
@liberal_party
that didn't do the due diligence in background checking of a prospective politician. How many other skeleton are in the closet?
There are many summer internship positions at
@VectorInst
!
Deadline: Jan 9th.
If you are interested in working on theoretical/algorithmic aspects of
#ReinforcementLearning
with me, please apply. I intend to recruit one or two interns.
Blog: Is Your Neural Network at Risk? The Pitfall of Adaptive Gradient Optimizers
Summary: Models trained using SGD exhibit significantly higher robustness to input perturbations than those trained via adaptive gradient methods such as Adam or RMSProp.
One of the reasons you see many young Iranians apply to your universities is that they are escaping a “condition”.
A “condition” that arrests those who protest against the deteriorating economy, tortures them to confess, and sentences them to execution.
#StopExecutionsInIran
I use Wikipedia everyday, not only for satisfying my curiosity, but also for my research. Yes, I read technical papers and books too, but in many cases Wikipedia is all that I need.
So consider donating to
@Wikipedia
, if you can!
#iloveWikipedia
Learning that DeepMind decided to close its Edmonton's office is upsetting. I don't know the detail, but I hope my friends and colleagues aren't unpleasantly affected.
If you are soon graduating from a PhD program and want to conduct world-class research in Machine Learning (including DL, RL, CV, NLP, etc), apply to become a Vector Postdoctoral Fellow!
@VectorInst
Deadline: March 31st
#PostdocPosition
In your opinion, what is the most interesting area of AI/ML that only a handful of researchers are currently working on?
Asking for a friend! (;
#MachineLearning
#FarFromMaddingCrowd
I wish I used this visualization last week in my ML course!
Mutual Information quantifies the dependency of two random variable. It's zero only when the r.v.s are independent. But covariance is only about their linear relation and can be zero even if the r.v.s are dependent.
Dimitri Bertsekas is the master of dynamic programming and RL. I've learned a lot from his Neuro-dynamic Programming (with J. Tsitsiklis) and Stochastic Optimal Control (with S. Shreve), and more recently the Abstract DP. I keep learning from them and using them in my own course.
My Reinforcement Learning course at ASU has started:
Class notes, video lectures and slides will be posted weekly at
First videolecture can be seen at
Dimitri Bertsekas
We have several postdoc positions at the Vector Institute. If you are a rising star in
#MachineLearning
, we want you to be here!
The deadline for this round is June 12th. After this, we have another round in September/October.
Tong Zhang has a new book on Learning Theory. I was introduced to it by
@CsabaSzepesvari
just today, so obviously I haven't read it yet. From a quick skim, it looks promising. Post it for now, so I don't forget.
#MachineLearning
#Statistics
If you are interested in doing a postdoc on
#ReinforcementLearning
, please apply to
@VectorInst
Postdoctoral Fellowship Program. I am looking for a theory-inclined person to start from around Summer 2023 or soon after.
Deadline: Feb 28th, 12AM
Please RT!
A good measure of diversity of an organization is the diversity of the senior management, high-rank officials, and people in position of power.
The apparent diversity of low-rank workers is probably a better indicator of broad exploitation than diversity.
If you are at
#NeurIPS2019
and curious to know the relation between distributional reinforcement learning and Fourier space, come to my poster today (Tue) at 10:45 AM, East Exhibition Hall B + C, poster No. 207.
A difference between an expert and a novice is in their robustness to uncertainty.
Someone who just learned something might understand what they learned very well, but they easily get confused if the presentation of the material slightly changes. Their understanding is fragile.
[This applies to Iranians] I am almost convinced that either you are with us or with them. There might have been a grey area 20 years ago where reform and dialogue could be hoped, but no more: it is Black or White now!
#MahsaAmini
#OpIran
"Intelligence is the computational part of the ability to predict and control a stream of experience" - Rich Sutton from
@UAlberta
@AmiiThinks
@DeepMind
at
@VectorInst
today.
I like this definition of intelligence (as well as William James and John McCarthy's).
Call it by its name: an RL architecture -- the cherry on top of the cake! 😉
BTW, the concept of World Model isn't new. It's a rebranding of the model in model-based RL (early 90s) and even earlier in Self-Tuning Regulators in adaptive control (early 70s).
#ReinforcementLearning
Chief AI Scientist Yann LeCun (
@ylecun
) is sketching an alternate vision for building human-level AI. LeCun proposes that the ability to learn “world models” — internal models of how the world works — may be the key. Learn more:
Like most sentient beings, I get stressed about the fast pace of machine learning research and the sheer volume of papers, which suffuse all my feeds.
To be far from this madding crowd, I visited a local math department’s library.
Preventing ordinary Iranians from being involved in scientific activies because of their country of origin/residence is disheartening and discriminatory, no matter how legally justifiable it is.
I am really sad that
@neuromatch
had to kick out all the Iranian students and TAs from it academy due to US sanctions! Politics should not be a barrier against scientific exchange and collaboration. Maybe they shouldn’t have registered it in a country with such crazy laws!
I feel shocked, sad, and sorry for my Afghan sisters and brothers who will live under the oppressive regime of Taliban. It is very scary to live under the regime following the 1400 year-old raw and literal interpretation of Islamic sharia.
#Afghanistan
The challenge of work-life balance is a false dichotomy to begin with. Work is not something outside life. It is a part of life itself. The real issue is about life-life balance.
The SARSA algorithm’s name comes from it usages of the State, Action, Reward, next State, and next Action in order to update the action-value function.
Based on this naming convention, the Q-Learning algorithm could be called SARS!
I am glad we don't!
#ReinforcementLearning
If you are an Area Chair for
#ICML2023
, please send a reminder to the reviewers to participate in discussions.
Reminding worked: there wasn't much reviewer activities the whole week until I sent a reminder last night.
If you are a reviewer, please read the rebuttals and react!
How the government of Ontario shoots itself in the foot by cutting the funding on
#AI
research at
@VectorInst
@CIFAR_News
.
@fordnation
@ToddSmithPC
Disclaimer: To be clear I am only representing my own opinion here, and not Vector's or CIFAR's.
A year ago today Mahsa Amini was killed by the thugs of the Islamic Republic of Iran. She was not the first and has not been the last to be demised by the IR in the past 44 years, but she was the butterfly that fundamentally changed most Iranians.
#MahsaAmini
#WomanLifeFreedom
I am teaching an Introduction to Machine Learning (grad-level) course this semester, so I thought I would brainstorm with my tweeps.
What topic do you wish was included in your own Intro to ML course that wasn't?
OR
What topic do you think should be included that is often not?
I will be at
#NeurIPS2018
.
If you are interested in reinforcement learning or want to know more about available positions at
@VectorInst
(research scientist or postdoc), talk to me.
A few days ago, the morality policy of the Islamic Republic of Iran murdered a 22 year woman called
#Mahsa_Amini
. You can see in this video how cruely they treat women (not Mahsa).
Trigger Warning: Watching this is upsetting.
Do you really want to know how Iranian morality police killed Mahsa Amini 22 year old woman? Watch this video and do not allow anyone to normalize compulsory hijab and morality police.
The Handmaid's Tale by
@MargaretAtwood
is not a fiction for us Iranian women. It’s a reality.
Very interesting!
My dream dynamical system to control, however, is the whole body of an Octopus!
I wish I knew more mechanics to model it, though I suspect a bit more wouldn't be enough – it needs a complicated nonlinear PDE, I believe.
Anyone interested?
We are releasing OstrichRL 🎉
The repository contains a musculoskeletal model of an ostrich in MuJoCo, a set of dm_control tasks for reinforcement learning, and motion capture data.
GitHub:
Paper:
Great advice, and applicable to other fields of knowledge too.
When reading a paper, I occasionally try to solve their problem before reading their solution. This is very time consuming and I often can't find a good solution, but it opens up my mind.
#DeliberatePractice
How to use the research literature. When approaching a difficult mathematical problem, first think deeply upon it on your own, using all the ideas you can muster, before consulting the work of others. Push your own ideas as hard as you can first, and read only afterward.
If you are interested in model-based reinforcement learning (MBRL), you want to read Iterative Value-Aware Model Learning, which is accepted at
#NeurIPS2018
.
So sad and unfortunate is
#IranPlaneCrash
. I am still in shock, and I guess I will be for a while, that how all those lives have perished so suddenly and needlessly.
I am very excited that I have joined the Vector Institute! I am looking forward to start new collaborations and build a strong research team there.
@VectorInst
We're very excited for you to meet Amir-massoud Farahmand, our newest Faculty Member who specializes in reinforcement learning!
Read what he has to say about his decision to pursue his research
@VectorInst
in Toronto:
@SoloGen
Interesting note by
@ccanonne_
on comparing KL vs TV.
We have the Pinsker inequality that says TV(p,q) < sqrt{KL(p||q)} (I am ignoring constants).
The trouble is that the RHS can become larger than one, while the LHS is at most one.
@vvanirudh
I spend ~4-6h (and sometimes even more) on each conference paper. This means that I have to dedicate most days of a week or two to reviewing for a conference.
I see this as a community service. I also learn those papers relatively well, as an added bonus.
Done with all my reviewing and area chairing responsibilities for 2019!
I think I should write about it, especially about the area chairing part. Maybe a blog post in the near future?
Are you a reviewer who cannot submit your report by the deadline?
Contact your senior program committee. Being a bit late might be OK. AWOL is not.
SPC is wondering whether they should wait for you or find an emergency reviewer. Don't leave them in the dark. Communicate!
I wish
#NIPS2018
didn't take a month to assign papers to reviewers and have them review papers in only three weeks, one of which is overlapped with the other major machine learning conference
#ICML2018
.
[late night geeky rant]
You want to design a risk-sensitive model-based
#ReinforcementLearning
agent. What is the right way to learn the model? MLE? Value-aware?
Tyler Kastner,
@MuratAErdogdu
, and I investigate this!
Paper:
Poster:
#NeurIPS2023
🧵000/111
Q: Does assassinating another country’s top general de-escalate the tension in the Middle East?
A: No!
I am worried that many innocent lives would be lost.
Deep Learning and Reinforcement Learning Summer Schools are now accepting applicants. The schools are in Toronto (July 25-August 3), and both have great lineup of speakers.
Apply and spread the news!
The deadline is March 26.
Old age has many symptoms. When I was a student, I used to feel delighted if the conference deadlines got extended. Nowadays I get annoyed. “I don’t want to work on this damned thing anymore”.
@ethanCaballero
They weren't considered clowns at all! They were all respected researchers. They were even pretty famous, though of course didn't have their current hyper-famous/god-like status.
It's true that (D)NNs weren't mainstream in 2000s, and many were skeptical of their importance.
Dimensionality Reduction for Representing the Knowledge of Probabilistic Models is accepted at
#ICLR2019
. This is by Marc Law, Jake Snell (
@jake_snell
), Richard Zemel, Raquel Urtasun, and myself.
The US is shooting itself in its brain again!
(1) I like many things about the US and its people. I enjoyed living there very much. But I find its government's malicious decisions to degrade human dignity repugnant.
(2) I feel sympathy towards students there.
#StudentBan
This is bad. ICE just told students here on student visas that if their school is going online-only this fall, the students must depart the United States and cannot remain through the fall semester.
Between MDPs and semi-MDPs: A framework for temporal abstraction in reinforcement learning by Richard Sutton, Doina Precup, and Satinder Singh (1999) has recently won the Classic Artificial Intelligence Journal (AIJ) paper award.
I gave a talk
@Carleton_U
on "Reinforcement Learning for High-Dimensional Problems: From PDE Control to Model Learning" and had a lot of interesting conversations.
Thanks to
@KomeiliMJ
for hosting me at
@CU_DataScience
.
Adaptive Agents Lab (yes, it is us!) had the pleasure of having Pierluca D'Oro
@proceduralia
this Monday to give a good talk about "Towards an Empirical Science of Neural Networks for Sequential Decision-Making".
#ReinforcementLearning
1/16 🧵
Submit your work to the Decision Awareness in
#ReinforcementLearning
Workshop at
@icmlconf
.
Decision awareness is a design principle: each module of an RL system should be trained to explicitly consider how its interaction with other modules affects the agent's performance.
Happy to announce the Decision Awareness in Reinforcement Learning workshop at
@icmlconf
!
We welcome contributions until May 27, 2022 AoE
The workshop website:
Hope to see you in Baltimore or virtually in July!
1/4
But then, "why aren't there only one or two books about this topic, but rows and rows of them? How much do they know that I don’t?"
My ignorance of the knowledge of the past made me depressed. I picked a book, and quietly left the library. The crowd started to rumble again.
Did you know that the optimizer significantly affects the robustness of NN? And Adam is the wrong answer!😯
"Understanding the robustness difference between SGD and adaptive gradient methods” dives deep into this.
Paper:
Code:
🧵1/4
"To make a remarkable contribution, you need to start by asking the right question. Good questions are:
1) tractable and fruitful.
2) secret. Questions that everyone are having are proportionally worthless."
[edited]
So Schmidhuber created neural networks in his own image.
And He blessed them, and said unto them, Be fruitful, and multiply, and replenish the artificial intelligence, and subdue it.
The most cited neural nets all build on our work: LSTM. ResNet (open-gated Highway Net). AlexNet & VGG (like our DanNet). GAN (an instance of our Artificial Curiosity). Linear Transformers (like our Fast Weight Programmers).
You keep seeing the same invited speakers all the time?
The rich get richer phenomenon for invited speakers has got exacerbated in the era of virtual talks. Whereas the travel cost (time&money) previously regularized how often one can present, that cost is now almost eliminated.
Texas abortion law shows that Artificial Intelligence and Machine Learning (or technology, more generally) is not the solution to even the most basic human needs, e.g., dignity and having control over one’s body.
You may ask what the relation of these two is?
Exactly my point!
I heard about the earthquakes in Turkey and Syria last night, but I'm just realizing how devastating they have been. Feeling sad for all the lost lives & wishing the safety of my friends and their families.
Is there any way to help?
#TurkeyEarthquake
#syriaearthquake
Congratulations Claas!
You did a very good job today. You are now officially my third PhD Candidate ever. I am looking forward to your PhD Graduate title.
I am taking a break from shitposting to announce that I have (if the protocol is submitted and I didn’t forget something) become a PhD candidate today. Thanks to my amazing supervisors and committee
@SoloGen
@igilitschenski
and
@WilCunningham
!
Let's not call this "Iran's attack", but "Islamic Republic's attack".
My estimate is that 80-90% of Iranians are against IR and an an overwhelming majority do not support IR's international politics.
I just met with my national security team for an update on Iran’s attacks against Israel. Our commitment to Israel’s security against threats from Iran and its proxies is ironclad.
Edison had 1000+ patents. He was a supervisor of a very productive R&D lab.
That is extremely valuable, but let’s not confuse it with being a lone genius or an exceptional innovator.
The same can be said about many other hyper-prolific creators.