I am very excited to announce
@ChandarLab
's fifth annual research symposium! Please join us on August 19th and 20th to learn more about our recent progress in deep learning, reinforcement learning, NLP, continual learning, optimization, and AI for science!
2 days of great talks
I am very excited to release this primer on lifelong supervised learning: . Lifelong learning is one of the most promising learning paradigms to achieve artificial general intelligence. 1/n
I am very excited to release the recordings of my Machine Learning lectures! You can watch the first lecture here: . If you want to follow the course, readings, lecture notes, and assignments will be made available at .
I am recruiting MSc and PhD students for Fall 2024. Application deadline is December 01. Please apply through the
@Mila_Quebec
supervision request process here: . More details about the recruitment process here:
I am very excited to release the recordings of my Reinforcement Learning lectures! You can watch the first-week lectures here: . If you want to follow the course, readings, lecture notes, and assignments will be made available at
I am happy to share that I will be joining
@polymtl
and
@MILAMontreal
as an Assistant Professor starting mid-Fall 2019! Looking for students to join my group and work on recurrent neural networks, lifelong learning, reinforcement learning, and NLP. 1/2
I am happy and proud to officially introduce my research group through this new lab website: . If you are interested in joining us,
@MILAMontreal
MSc/PhD applications are now open! Deadline: Dec 15. For more info:
I am happy to announce that I have been promoted to Associate Professor with tenure at
@polymtl
. This is an achievement for the entire
@chandarlab
! I want to thank my awesome students, without whom this would not have been possible! 1/n
Join my lab! I am currently recruiting new students for admission in the Fall of 2024. If you're interested in working with me, submit your application via Mila's supervision application process. To find out more, click here:
@Mila_Quebec
I am teaching a new graduate-level Machine Learning course
@polymtl
/
@MILAMontreal
in Fall 2020 - . The course is online. If you are a student in any of the Quebec Universities, you can register for the course here:
I am looking for Masters and Ph.D. students to start in Fall 2020. If you are interested in working with me at
@MILAMontreal
(), the deadline is Dec 15. More info here: . You can also come to talk to me at
#NeurIPS2019
.
I have multiple open MSc/PhD positions on memory augmented neural nets, RL, Lifelong Learning, NLP for Fall 2022 at
@ChandarLab
/
@Mila_Quebec
/
@polymtl
!
Details:
Applications due Dec 1st:
I am truly honored to be named a Canada Research Chair (CRC) in Lifelong Machine Learning in addition to my already existing Canada CIFAR AI Chair by
@CIFAR_News
! This is a recognition to all the exceptional work of my students
@ChandarLab
! Thank you
@NSERC_CRSNG
!
Congratulations to our Core Academic Member Sarath Chandar, who was appointed Canada Research Chair in Lifelong Machine Learning this week and received a grant from the Canadian government to further his research work.
Read the full annoucement :
I will be at
#NeurIPS2023
from 14-17. I am looking to recruit MSc/Phds to work on
1. Optimization for DL and RL.
2. Continual/Lifelong Learning.
3. LLMs. (eps. bias, interpretability)
4. ML/RL for scientific discovery.
Email me to meet there if you have relevant background!
I am teaching a course on scientific and technical communication for CS PhD students this Fall. Please suggest great books, articles, blogs posts on how to read, write, review technical papers, and also how to present your research in an effective way.
I noticed a footnote in this paper. Why the rush?
It is strange to arXiv a paper with a "known* bug, but the bug is mentioned in a tiny footnote. This could be fixed by re-training in a few days. We as a community, need to slow down.
🎉 Excited to announce our work "Recall to Imagine (R2I)" got an oral spot at
#ICLR2024
! R2I is a generalist and computationally efficient agent that shines in RL memory tasks, showing superhuman performance in most complex memory tasks. Kudos to
@M_R_Samsami
&
@artemZholus
! 1/n
We introduce a new regularizer for CNNs called PatchUP. PatchUP alleviates some of the drawbacks of the previous mixing-based regularizers - . Work led by Mojtaba Faramarzi. We also opensource our code to reproduce the results here: .
FINAL UPDATE: On June 24th, Armando Solar-Lezama (Professor in EECS and COO/Associate Director of CSAIL, MIT), Tonio Buonassisi (Professor of Mechanical Engineering, MIT), and Yoon Kim (Assistant Professor in EECS and CSAIL, MIT) released a public statement regarding the paper.
Doina Precup,
@rpascanu
and I are very excited to announce the first International Conference on Lifelong Learning Agents (CoLLA)!
If you are working in lifelong/continual learning, meta-learning, multi-task learning, or OOD generalization, consider submitting your work to CoLLA!
At
@ChandarLab
, we are happy to announce the launch of our assistantce program to provide feedback for members of communities underrepresented in AI who wants to apply to high-profile graduate programs. Want feedback? Details: . Deadline: Nov 15!
Time to stop stereotyping India like this! Just look around. All your major companies (including this Twitter) are run by Indians and that should give you some clue.
Inspired by
@kchonyc
, I will be doing free-of-charge pro-bono office hours to talk with organizations that have the need for Artificial Intelligence (
#AI
) and Machine Learning (
#ML
) but does not have in-house expertise in AI/ML. For more details:
Are you tired of manually creating new tasks for Lifelong RL? We introduce Lifelong Hanabi in which every task is coordinating with a partner that's an expert player of Hanabi. Work led by
@HadiNekoei
and
@akileshbadri
.
paper:
@Mila_Quebec
1/n
Our work on training RNNs for lifelong learning got accepted at Neural Computation! Joint work with
@shagunsodhani
and Yoshua Bengio. The paper discusses how can we tackle both catastrophic forgetting and capacity saturation jointly.
If someone wants to do AI research in India, this is definitely the best place to do so! I’m looking forward to all the great work that is going to come out of Wadhwani School of Data Science and AI..
I am very proud and happy to announce that our
@ChandarLab
MSc graduate Ali Rahimi received the best Master's thesis award from the Canadian AI Association for 2024!
Ali's masters thesis shows that SOTA MBRL methods like Dreamer and MuZero are not adaptive and he also has a fix!
We are very pleased to announce Ali Rahimi Kalahroudi (Université de Montréal) as the recipient of the CAIAC 2024 Best Master's Thesis Award. Ali's thesis was "Towards Adaptive Deep Model-Based Reinforcement Learning."
Happy to announce that the 4th edition of the lifelong learning workshop will happen at
#ICML2020
@icmlconf
.
We (
@shagunsodhani
,
@ravi_iitm
, Doina Precup, and I) are experimenting with a new (or very old?) format for the workshop this year. Stay tuned for more details and CfP!
In this recent interview with
@CIFAR_News
, I talk about lifelong machine learning, my experience with CIFAR DLRL Summer Schools, and how great is Canada for collaborative research :)
Looking forward to the
#DLRL
panel tomorrow!
@CoLLAs_Conf
#CoLLAs2022
was a big success! We had 64 accepted papers, 230+ participants, 8 exciting invited talks! This is a thank you thread for everyone involved: 1/n
The 14th **Asian Conference on Machine Learning** is coming to India! It will be in a hybrid mode, offline meeting in Hyderabad.
Submit your works here!
There are both journal (ddl 26th May)and conference (ddl 23 June) tracks..
Check for more details!
CoLLAs 2022 CFP is out! Highlights: Proceedings in PMLR, revise-resubmit option, 9 pages + an optional 10th page to avoid wasting hours fitting your paper to a strict page limit, appendix in same pdf.
Foundation models are pure marketing and unfortunately good marketing always wins. We don’t use it in my group. We just call them as large language models.
Lifelong Hanabi is now accepted at ICML 2021! If you want to try your new lifelong learning algorithms in Hanabi, here is the code for the benchmark:
#ICML2021
@icmlconf
@Mila_Quebec
Are you tired of manually creating new tasks for Lifelong RL? We introduce Lifelong Hanabi in which every task is coordinating with a partner that's an expert player of Hanabi. Work led by
@HadiNekoei
and
@akileshbadri
.
paper:
@Mila_Quebec
1/n
Introducing an improved adaptive optimizer: Adam with Critical momenta (Adam+CM)! Unlike traditional Adam, it promotes exploration that paves the way to flatter minima and leads to better generalization. Link to our paper:
Work led by:
@pranshumalviya8
I will be teaching a graduate-level Machine Learning course
@polymtl
/
@Mila_Quebec
in Fall 2021 - . The course is online. If you are a student in any of the Quebec Universities, you can register for the course here:
We are very excited to release RLHive (), an RL research framework in PyTorch. RLHive supports both single-agent and multi-agent RL!
This is an outcome of a year-long effort by
@dapatil211
,
@HadiNekoei
,
@alirkay
,
@saikrishna_gvs
.
After several months of hard work, I am proud to share the release of RLHive, a research framework for RL in PyTorch for both Single Agent and Multi Agent use cases.
Github:
Docs+Tutorials:
1/n
The Computer and Software Engineering Dept. at Polytechnique Montreal is hiring for a tenure-track faculty position in Trustworthy AI!
Eligible candidates will be nominated to become affiliated with
@Mila_Quebec
!
Link to the official call in the next tweet!
Deadline: 8 April
If you are interested in lifelong learning / continual learning, please check this thread for the amazing list of speakers and panellists at our 4th Lifelong Learning workshop
@icmlconf
this Saturday!
You can ask your questions to the panellists here:
Our lab has its own Twitter handle now! Managed by the awesome students :) Please follow us to hear more about our new research projects. We will also tweet about the works of others that we are excited about!
If you wanna learn more about the recent advances in deep learning, reinforcement learning, and NLP that has come out of my lab in the past year, consider attending our lab's annual research symposium on Aug 8 and 9:
You can join remotely too!
Reflecting on this a bit further -- I think this episode really emblematic of how broken evaluation is in our field right now. In general, data work is important, and creating benchmarks is work that has been historically undervalued. However:
Want to know more about what is happening at
@ChandarLab
? Please join our annual research symposium () virtually or in person (Montreal) this August 11! You will hear my students talking about lifelong learning, reinforcement learning, NLP, and DL!
Geoffrey Hinton is right. So-called open sourcing of the biggest models is completely crazy.
As AI models become more capable they should become increasingly useful in bioweapons production and for use in large-scale cyber attacks that could cripple critical infrastructure.
The first edition of the Conference on Lifelong Learning Agents (
@CoLLAs_Conf
) will happen in hybrid mode in Montreal from Aug 22 - Aug 24! Registration for virtual and in-person attendance is now open! .
My department at
@polymtl
is hiring a tenure-track professor in computer and software engineering. Strong ML candidates can get
@Mila_Quebec
affiliation. If you are working at the intersection of Systems and ML, we are looking for you! Deadline: April 05.
Do you work in lifelong/continual/meta/in-context/transfer/multi-task/few-shot/zero-shot/curriculum/OOD/active/online learning paradigms in ML/RL?
CoLLAs is the premier venue for these topics! Submit your best work to CoLLAs 2024!
Deadline: 15 Feb 2024
#CoLLAs2024
@CoLLAs_Conf
Happy to announce that the 4th edition of the lifelong learning workshop will happen at
#ICML2020
@icmlconf
.
We (
@shagunsodhani
,
@ravi_iitm
, Doina Precup, and I) are experimenting with a new (or very old?) format for the workshop this year. Stay tuned for more details and CfP!
This is a great list of 64 interesting papers focusing on lifelong learning, meta learning, few shot learning, transfer learning, reinforcement learning! Please do check the papers and their video presentations!
@CoLLAs_Conf
You can find all the 64 accepted papers (pdf)
@CoLLAs_Conf
#CoLLAs2022
and their video presentations here: . You can watch these video presentations even if you have not registered to attend the conference!
Of course a sad day in history! It is important to commit to make education and research opportunities more equitable. To that, I will keep open exclusive internship positions for 2 undergraduate
#WomenInAI
every year starting Summer '21 at
#ChandarLab
@Mila_Quebec
@polymtl
.
Frustration after staying in a few hotels for a vacation - all North American 4-star hotels should go learn about hospitality and customer service from Indian 4-star hotels. Indian hotels provide 50X better service for 5X lower price.
Probing the representations of generative dialogue models shows interesting trends. We observe that probing tasks can holistically evaluate generative dialogue models. The full paper can be found here ! Work led by
@prasannapartha
.
#NLProc
1/5
Lecture
#6
of the Machine Learning course is now available in the playlist:
. This lecture discusses about empirical risk minimization, bias-variance trade-off, and different approaches for classification.
#MachineLearning
If you have a recent
@TmlrOrg
paper on lifelong/continual/multi-task/meta/transfer/online/active/ood learning, consider presenting the paper as a poster at
@CoLLAs_Conf
and get a CoLLAs certification by TMLR! I’m very excited about this collaboration with
@TmlrOrg
!
We are excited to announce a new "Journal Track" at CoLLAs 2023! This track gives an opportunity for recently published journal papers (that are aligned with the theme of CoLLAs) to be presented as posters at CoLLAs! Details here:
1/n
Lecture
#4
of the Machine Learning course is now available in the playlist:
. The first part introduces statistical decision theory and the second part covers basis function regression models and geometric view of linear regression.
#MachineLearning
Check out our recent work on a new experimental setup and metric for evaluating model-based
#RL
methods. Work led by
@harmvanseijen
(Microsoft Research) and my amazing interns
@HadiNekoei
and
@evanracah
.
We introduce a new experimental setup and metric for evaluating model-based RL methods. The metric measures how quickly an RL methods adapts to local changes in the environment.
Jana was the first postdoc
@ChandarLab
and I’m extremely proud that he is starting his own lab
@DalhousieU
! Jana is a great researcher, excellent mentor, and he truly cares about training students to do impactful research! I strongly encourage students to apply to his lab!
Pleased to share that I will be joining
@DalhousieU
as an Assistant Professor in Jan '24. :) I have openings for Master’s & PhD positions in Deep Reinforcement Learning. Details: .
Please share with potential candidates and within your network. Thanks!
🎉 Excited to announce our work "Recall to Imagine (R2I)" got an oral spot at
#ICLR2024
! R2I is a generalist and computationally efficient agent that shines in RL memory tasks, showing superhuman performance in most complex memory tasks. Kudos to
@M_R_Samsami
&
@artemZholus
! 1/n
Lecture
#3
of the Machine Learning course is now available in the playlist:
. In this lecture, we discuss ways to avoid overfitting, basic ML pipeline, and K-NN for classification.
#MachineLearning
Why should we consider GPT-4 as state of the art in research? It’s a black box product. OpenAI and Microsoft can write papers on GPT-4 because they are selling the product. We should be using only open source LLM models for research.
CoLLAs is the community of researchers who truly care about the continual learning problem! If you work in any non-IID / non-stationary learning setting, CoLLAs will be your home conference! Consider joining us for CoLLAs 2024!
Do you work in lifelong/continual/meta/in-context/transfer/multi-task/few-shot/zero-shot/curriculum/OOD/active/online learning paradigms in ML/RL?
CoLLAs is the premier venue for these topics! Submit your best work to CoLLAs 2024!
Deadline: 15 Feb 2024
#CoLLAs2024
@CoLLAs_Conf
If you wanna learn about the whole landscape of efficient Transformers, I highly recommend this excellent suvery on faster and lighter Transformers by my postdoc Quentin Fournier: (Accepted in ACM Computer Surveys)
Canada is looking for talented immigrants and you can get your PR in a year and citizenship in less than 5 years (outside Quebec). These H1B visa holders should just consider moving to Toronto :)
Stories below of Indian H-1B visa holders jumping from big tech (Google & Apple) to startups (Coinbase & Peloton), getting laid off then having to leave the country because they couldn’t find jobs in 60 days are harrowing.
It’s why H-1Bs end up feeling like indentured servants.
Only one week left for the application deadline! In addition to the listed topics (memory augmented networks, learning through language interaction, optimization, lifelong learning, RL), I am also looking for MSc/PhD students to work in the intersection of ML and Drug Discovery.
I have multiple open MSc/PhD positions on memory augmented neural nets, RL, Lifelong Learning, NLP for Fall 2022 at
@ChandarLab
/
@Mila_Quebec
/
@polymtl
!
Details:
Applications due Dec 1st:
Can large language models consolidate world knowledge? The answer turns out to be "NO". I am very excited to present to you our
@emnlpmeeting
2023 paper (main track) which studies this important limitation of LLMs. Work led by my amazing PhD student
@GabrielePrato
!
The ability to combine information from different training documents is crucial for many applications. We introduce the first study on the knowledge consolidation capabilities of LLMs:
I will be at NeurIPS from 14-17 Dec. I will be recruiting a postdoc in Reinforcement Learning. Please email me to meet during the conference if you are looking for a postdoc position starting soon!
#NeurIPS2023
I wasted my time reading this article. Even a first year phd student would have the same vision. What is so special to write an article about a trivial vision which is not even published in arxiv yet? This fan following must stop. Doing PR even before the preprint is next level!
Please check out our recent preprint on maximum reward formulation in
#RL
: . This work improves upon our
#ICML2020
paper on RL for drug discovery (). Work led by
@saikrishna_gvs
.
A recent interview about our ongoing efforts on RL for drug discovery. Led by
@saikrishna_gvs
and
@boris_luv
of . Accepted at
@icmlconf
. The preprint can be found here: .
Can’t agree more! There is no good reason why a paper with Stanford-only author list gets all citations for the hard work of other research groups! Most people hate the name for this marketing only.
@chrmanning
@tdietterich
@roydanroy
@karpathy
@ylecun
@percyliang
@RishiBommasani
The jazz analogy is quite good from a credit assignment view. Very deeply, i think the disagreements are around other academics thinking "why do you get citations, credit and eventually prizes for consolidating other people's breakthroughs and putting a catchy term around it".
I feel like most courses/textbooks focus on the mechanical aspects of linear algebra. Anyone know of any resources to develop good intuitions about concepts?
It takes less than a year to get PR in Canada (except Quebec). Indians in America should consider immigrating to Canada! Unlike America, Canada is actually more welcoming for highly skilled immigrants!
Conference proceedings are equally junk and we can’t keep up with them too. Research is a noisy process and a researcher should be able to separate the signal from the noise. Peer review was never perfect and will never be perfect. Arxiv does more good than bad!
@tdietterich
@TaliaRinger
@mmitchell_ai
@ErikWhiting4
@arxiv
arXiv is a cancer that promotes the dissemination of junk "science" in a format that is indistinguishable from real publications. And promotes the hectic "can't keep up" + "anything older than 6 months is irrelevant" CS culture.
>>
I’m extremely proud of my student
@Megh1211
who has been awarded the Antidote scholarship for NLP! I received the Antidote scholarship during my phd at
@UMontrealDIRO
and
@Mila_Quebec
and very happy to see my student joining the list of recipients! 😃
🎉 Congratulations to Megh Vipul Thakkar, a Master's student (Polytechnique/Mila) who has been awarded the eight Antidote Scholarship from
@druide_info
in Natural Language Processing (NLP), worth $20,000. Congratulations!
Quality of reviews are already too bad in ML conferences! We should not invite inexperienced and new researchers to start reviewing. They need more experience doing research before starting to review others works.
I'm happy to nominate new reviewers for
#NeurIPS
. If you are interested, send me your email, OpenReview profile, and a short blurb about your research background and reviewing experience — inexperienced/new researchers are welcome too!
Every ethnicity has cheating and non-cheating members. Most asians I have seen are sincere, hard-working, smarter, talented. I think part of the reason for this prejudice is just the insecurity of non-asians who make such comments!
@thegautamkamath
Yes. Very sadly. I have been told directly in person that Asian tend to cheat, etc. multiple times. I fought back every single time. But that is also because I have certain privilege's that enabled me to do so, can't image how bad it is for those without that.