This Moroccan Arab Muslim first-generation woman just gave her first long talk for her award-winning paper at
#ICML2022
! I dedicate this achievement to all the underrepresented groups that I proudly represent!
So overwhelmed by all the support that I received! Many thanks! 1/2
I'm so grateful and honored to receive the Microsoft Research PhD Fellowship
@MSFTResearch
!!🥳
This fellowship means that my research and that of my group led by
@andrewgwils
is recognized as meaningful and impactful for the machine learning community! 1/3
I'm so proud that our paper on the marginal likelihood won the Outstanding Paper Award at
#ICML2022
!!! Congratulations to my amazing co-authors
@Pavel_Izmailov
,
@g_benton_
,
@micahgoldblum
,
@andrewgwils
🎉
Talk on Thursday, 2:10 pm, room 310
Poster 828 on Thursday, 6-8 pm, hall E
I'm happy that this paper will appear as a long oral at
#ICML2022
! It's the culmination of more than a decade of thinking about when the marginal likelihood does and doesn't make sense for model selection and hyper learning, and why. It was also a great collaborative effort.
We talk a lot about LLMs but you know what we don’t talk enough about? PhD student mental health, community building, and effective research practices!
That’s why Yaniv Yacoby and I are organizing a workshop at NYU entitle “Everything I wish I knew as I navigated my PhD”. 1/3 👇
I'm sorry for the negative attention I brought on
@LotfiSanae
in what should have been a joyous moment and who has authored an amazing & exciting paper on a complex topic. My intention was scientific integrity, but I lost sight of the broader context of this academic debate &
It was a great honor to be distinguished as a Rising Star in ML by
@ml_umd
🥳 Many thanks to the organizers and professors with whom I had great discussions!
Next stop:
#NeurIPS2023
; so excited to share our work on generalization bounds for LLMs and host
@MuslimsinML
there!
Went to Harvard to give a talk about our work on the marginal likelihood and PAC-Bayes bounds, ended up meeting
@ConanOBrien
there. It turns out he’s a big Bayesian!
Can LLMs generalize meaningfully beyond their training data?
We answer this question by computing the first non-vacuous generalization bounds for LLMs. Wanna learn more? Find us the M3L and SSL workshops 😉
w/
@m_finzi
@KuangYilun
@timrudner
@micahgoldblum
@andrewgwils
Had a great time presenting our work on PAC-Bayes bounds. Thanks everyone who stopped by; many insightful discussions!
I’m at
#NeurIPS2022
this week, ping me if you want to discuss generalization in deep learning or Taylor Swift concert tickets!
Paper:
🚨
#NeurIPS2022
poster today: 4-6pm, Hall J
#306
🚨
Why do CNNs generalize so much better than MLPs? Why can neural networks fit random labels and still generalize? What is the value of encoding invariances in our models?
1/N
We’re organizing Muslims in ML
@NeurIPSConf
this year 🥳
Submit a two-page abstract if you would like to showcase and present your work at the workshop by September 22 ⏰
Excited to announce that the Muslims in Machine Learning (MusIML) workshop is back
@NeurIPSConf
!
If you self-identify as Muslim, or work on research that address challenges faced by Muslims, we'd love to showcase your work. Submit an abstract by Sep 22:
This competition resulted in hundreds of submissions and many interesting solutions that will be presented by the winners 🏅 tomorrow starting 1pm ET 🎉 Join us to discover these scalable approximate inference methods for Bayesian Deep Learning:
I heard a rumour there is this amazing Approximate Inference in Bayesian Deep Learning competition at
#NeurIPS2021
tomorrow, starting at 1 pm ET. From what I understand, the winners will be revealing their solutions, and the link to join is . 🤫
Very excited to announce that I will be joining the Center for Data Science at New York University as a PhD student.
I am also very honored to receive the DeepMind Fellowship.
Thank you
@DeepMind
and
@NYUDataScience
!
Meet Sanae Lotfi, a PhD CDS student who is one of three DeepMind Fellows joining us this academic year. We’re also proud to note that Sanae has recently won the McKinsey First Generation Achievement Award earlier this month. Learn more on our blog:
So happy we ended up having a last minute talk by
@MarzyehGhassemi
on Ethical AI in Health
@MuslimsinML
!
I am a big fan of
@MarzyehGhassemi
’s work and I also like her so much as a female researcher, mentor, and role model. Very thankful I got a front seat to her talk! 🥳
Being a Visiting Researcher at Meta has been amazing, but it’s coming to an end!
I’m looking for internship opportunities for Summer 2024, possibly my last PhD internship. I’m interested in the foundations of deep learning with a recent focus on LLMs. Reach out if you see a fit!
That’s not true. We re-ran any experiments affected by a minor bug (one panel only in the main text) in the camera ready, and the qualitative results remain unchanged. We engaged extensively with all your questions via email, up through you publicizing your review.
Not sure if this paper deserves such an award... After pointing out a bug, the authors reran some experiments. But the results in the revisions contradict each other and the authors never replied to my Qs in the email thread again, so yay🔥
PC review:
Wrapped up my Boston visit with a talk at
@MIT_CSAIL
! Many thanks to all the brilliant researchers who invited me or made time to chat! Thank you
@MarzyehGhassemi
,
@irenetrampoline
and Yaniv for being amazing hosts as well!
Bonus: I discovered more Taylor Swift fans in academia!
It has been extremely stressful not to be able to defend our work on Twitter, just to avoid an unproductive discourse. We wanted to enjoy ICML but were instead stressed out all week glued to Twitter as our integrity was called into question. 11/N
I am at ICML this week and would love to meet and chat with some of you there! If you would like to talk about generalization in deep learning and other topics, ping me here or find me somewhere in the Baltimore Convention Center!
#ICML2022
Our extended paper on the marginal likelihood was accepted to JMLR 🎉
In this version, we expand on the marginal likelihood’s connection to PAC-Bayes bounds, its approximations, and its use for architecture search.
Check the thread👇for more details and stay tuned for more!
Last year at ICML, we presented marginal likelihood pathologies in model selection and hyper learning. We now have a 60 page JMLR extension featuring: 1) should we be comforted by connections with PAC-Bayes? 2) approximations; 3) architecture search.
1/16
Thrilled that this work with
@Pavel_Izmailov
,
@g_benton_
,
@micahgoldblum
, and
@andrewgwils
got accepted for a long oral presentation at ICML 2022. It was fun to investigate and articulate the subtle ways in which the marginal likelihood is different from generalization!
Come join me tomorrow, alongside
@polkirichenko
and
@shiorisagawa
, to discuss methods and challenges in tackling distribution shift in deep learning. No matter what you are working on, come share with us how distribution shift affects your work and/or its applications!
#ICML2022
We are excited to present our breakout session on robustness to distribution shift at
@WiMLworkshop
@icmlconf
together with
@shiorisagawa
@LotfiSanae
! Join our session & discussion tomorrow, Monday July 18, at 11am at the Exhibit Hall G at Level 100 Exhibition Halls 🙂
#ICML2022
Just know Andreas that none of us hold a grudge against you and hope we can interact in a friendly way going forward. Thanks for extending an olive branch! 12/N, N=12.
🔥Talk tomorrow, Dec. 5, at the
@NorthAfricansML
workshop, 10-11 AM GMT+1 🔥
"Are the Marginal Likelihood and PAC-Bayes Bounds the right proxies for Generalization?"
1/11
Big thanks to my amazing co-authors
@Pavel_Izmailov
,
@g_benton_
,
@micahgoldblum
, and my incredible advisor
@andrewgwils
who put a lot of effort into this work and supported me all along!
Want to learn more about this work? check our poster 828, hall E between 6 and 8 pm EST!
When is it responsible to deploy a general purpose AI system?
@OpenAI
’s
@_lamaahmad
is answering this question right now
@MuslimsinML
!
Very exciting talk about identifying the potential risks of
@OpenAI
’s systems through Red Teaming!
This work was a team effort. We all care about scientific rigor and stand by the work. I do not think it is acceptable to single out any other author and direct attacks onto them. We all take responsibility and are committed to addressing questions about it. 2/N
@biotin10000mcg
"A formal definition of a first-generation college student is a student whose parent(s) did not complete a four-year college or university degree." My parents actually never went to school. So proud of them for raising us and educating us to appreciate science this much!
I'd like to thank my incredible advisor
@andrewgwils
for his continuous support, my supportive colleagues
@Pavel_Izmailov
and
@micahgoldblum
for their feedback on my proposal, and my amazing masters advisor
@69alodi
and Prof. Julia Kempe for supporting my application! 2/3
Very cool work! I was just yesterday mentioning during the lab I TA that we have more intuition about the function space than the parameter space, and that it would be cool to thoroughly investigate function space MAP estimation — we don’t have to wonder anymore! 🥳
When training machine learning models, should we learn most likely parameters—or most likely functions?
We investigate this question in our
#NeurIPS2023
paper and made some fascinating observations!🚀
Paper:
w/
@ShikaiQiu
@psiyumm
@andrewgwils
🧵1/10
Some of the PhD struggle is necessary and a byproduct of growth, but some of it is not and needs to be dealt with effectively and compassionately! We’ll try to figure which is which with amazing peers
@SCSatCMU
👇
Let’s keep discussing PhD student mental health as a community!
📣📣
@SCSatCMU
PhD students: I'm so excited to share that
@charvvvv_
,
@LotfiSanae
, Yaniv Yacoby, and I are bringing this amazing workshop to CMU!
Join us to reflect on PhD student mental health, community-building🫂, and academic culture. (1/n) 👇
I just donated too! Thanks
@bneyshabur
for bringing this into my feed! So grateful and proud of
@ml_collective
and
@DeepIndaba
for their effort to support African researchers!
P.S: you can also help with your time through the Indaba Mentorship Programme:
8 researchers from Nigeria got accepted into Deep Learning Indaba
@DeepIndaba
but couldn't afford the trip. Access to research, mentorship, and networking opportunities like this is vital for early-stage researchers. You can help make their trips possible!
@shortstein
The reviewer in me likes it, the author in me does not. What a dilemma!
In general, I believe we should treat the papers we review the same way we want our papers to be treated, so I'll probably pass on this option whenever possible😅
I also don’t appreciate your suggestion that we did not deserve the award or that people should complain to ICML. We are proud of this work, and feel it is deserving of such support. 3/N
Big thanks to all of my collaborators, labmates and mentors who inspire every day and to my amazing partner for his ample feedback and help with my research proposal! 3/3
We explore many ways in which the question ``how likely is my prior to generate the training data?’’ differs from the question we care about ``how likely is my posterior to model withheld points from the same distribution?’’. We also distinguish this from hypothesis testing. 5/N
The workshop is happening on Monday April 3rd at NYU (60 5th Ave) from 12-2pm and It’s open to all NYU PhD students. You can find the link to register in your inbox or DM me to send it to you! 2/3
@BlackHC
That’s also incorrect. We did not cherry-pick the experimental setting as I used exactly the same models and checkpoints in both arxiv versions. I did not do any re-training after we fixed the bug. I’m sorry but we don’t think you’re engaging in good faith. Thanks for the review.
We are agnostic as to whether the CLML is better or not than validation, and this is not at all the focus of our work. Sometimes it will be, others it won’t. If one can do something simpler and achieve better performance, we are supportive of that. There is nothing to hide. 9/N
The marginal likelihood is regularly used as a proxy for generalization, in seminal early work, and in recent papers, for hyperparameter learning, and model selection. While it can sometimes be practical for these purposes… 4/N
So excited to be presenting our work on PAC-Bayes bounds 🔥tonight🔥 alongside an amazing lineup of female speakers!
Join us if you want to hear more about this work w/
@m_finzi
,
@snymkpr
, A. Potapczynski,
@micahgoldblum
,
@andrewgwils
; or to chat!
Paper:
Join us at the Women in AI Ignite today 6:00 pm CT
@NeurIPSConf
Great lineup
#womeninai
A room full of inspiring female figures reminds me of how important it is to support women around the world to live freely and reach their true potential.
#WomenLifeFreedom
@anoushnajarian
Been using CoLA pretty much since it came out and I like it a lot! It makes large matrix inversion, eigenspectrum decomposition and accessing other quantities I need for my research much more memory-efficient and scalable. Works with PyTorch, JAX and supports GPUs. Try it out!
Like differentiation, numerical linear algebra is a notorious bottleneck that is ripe for automation and acceleration. At
#NeurIPS2023
we introduce Compositional Linear Algebra (CoLA)!🥤
w/ A. Potap, G. Pleiss,
@andrewgwils
. 🧵 [1/7]
I’m so excited to be part of this workshop and host Yaniv Yacoby who created several successful initiatives around improving PhD student mental health and community at Harvard.
I hope this workshop results in follow-up initiatives to strengthen our graduate community at NYU! 3/3
Big thanks to the
@NorthAfricansML
workshop organizers as well. It's great to see a NeurIPS workshop that is fully accessible in the North African timezone. Hopefully we get to see NeurIPS in Africa soon!
11/11
@FelixHill84
@BlackHC
I have no doubt! I realize how Twitter can make things escalate quickly and amplify misunderstandings, especially when it comes to technical and nuanced discussions!
The final phase of our
#NeurIPS
2021 competition "Approximate Inference in Bayesian Deep Learning" has officially started!
You can test the fidelity of your approximate inference procedure with the opportunity to present it at the
#NeurIPS
Bayesian deep learning workshop (+$)!
In short, the LML has already been compared extensively to various types of validation (Minka on automatic PCA, Rasmussen for GPs..), and in some cases is preferable, in others it is not. As we show, the CLML will typically be more aligned with generalization than the LML. 8/N
We did not remove the comparison to non-BMA cross validation from the CR. We just never had it in arxiv v1 and we built the CR on v1 and not v2. There was no deliberate action. We have no issue with BMA validation doing better for architecture search & will highlight in v3. 10/N
Your main concern at this point seems to be whether the CLML is competitive with validation loss. We have already largely responded to this point. See the attached reply we had in your blog. 7/N
Although it is not the focus of our paper, we show how trivially modifying the log marginal likelihood (LML), for the CLML, can lead to a better correlation with generalization for model comparison and hyper learning. We stand by this observation. 6/N
With these bounds we find that models which encode locality and equivariances are more compressible when these biases match the structure in the dataset, and hence have tighter generalization bounds.
6/N
@nsaphra
@BlackHC
@hewal_oscar
Thanks for pointing that out Naomi. The camera ready has updated figures indeed and a discussion about how we never considered cross validation to be a competing metric (see appendix L).
@69alodi
@NYUDataScience
@polymtl
@DS4DM
It was a pleasure to be supervised by you and Dominique. I am so grateful to you,
@69alodi
for all the freedom and guidance you offered. I also feel immense gratitude towards our team at
@DS4DM
. Congratulations on creating this positive, supportive, and collaborative environment.
Join the Q&A session at 18:30 ET to learn more about our work "Stochastic Damped L-BFGS with Controlled Norm of the Hessian Approximation", with Tiphaine Bonniot, Dominique Orban and
@69alodi
OPT2020 Workshop schedule:
Paper:
Our bounds also explain why transfer learning works; the difference between a pre-trained checkpoint and the fine-tuned model is far more compressible than the difference between a random initialization and the corresponding model trained from scratch.
8/N
Typical transfer learning pipelines involve initializing at pre-trained weights and hoping that relevant learned information magically transfers even when the weights change during fine-tuning. But you can transfer so much more than just initialization! 1/4
On the other hand, when these structures are broken, the models are less compressible and hence generalize worse even than generic MLPs. We also find that checkpoints which fit random labels are not compressible, explaining why such flexible models generalize well.
7/N
By training in a low-dimensional affine subspace and using aggressive quantization, we design a practical method to achieve the trade-off between the training error and the KL divergence as we change the subspace dimension.
5/N
PAC-Bayes bounds are another expression of Occam’s razor where simpler descriptions of the data generalize better, that can be used to understand generalization in deep learning.
5/11
🚨
#NeurIPS2022
poster today: 4-6pm, Hall J
#306
🚨
Why do CNNs generalize so much better than MLPs? Why can neural networks fit random labels and still generalize? What is the value of encoding invariances in our models?
1/N
Work by Germain et al., 2017 () has shown that minimizing the PAC-Bayes bound is equivalent to maximizing the marginal likelihood when choosing the loss function to be negative log-likelihood.
8/11
@micahgoldblum
You should at least add her in the acknowledgement paragraph for moral support and feedback (pretty sure she gives you feedback, doesn't she?)
Previous works choose between loose data-independent bounds and data-dependent priors which poorly explain generalization. Instead, data-dependent bounds essentially say that neural nets generalize because they have good validation accuracy …
3/N
@69alodi
I am very lucky that I have and had your support as my master's co-supervisor and my mentor to this day! The trust, support and mentorship you keep offering me are so much appreciated! Thank you very much Andrea!