Assistant professor at Bar Ilan University. Primary research is Machine learning, computational biology, and signal processing.
Israel
Joined February 2018
- Tweets4.7K
- Following3.4K
- Followers1.9K
- Likes46.3K
At this point I'm not surprised that most reviews I got are 100% ai generated (by pangram), I'm just disappointed that some of them are based on shitty models.
#ICLR2027 @iclr_conf
1/ ML is shifting towards “foundation models are all you need.”
These huge networks can do many things, but even fine-tuning now demands unreasonable amounts of GPU memory.
How can parameter-efficient fine-tuning (PEFT) help?
A thread 🪢
4/ Flatten → SUMO (NeurIPS 2025): is a Muon-style orthogonalization, computed exactly inside a low-rank subspace. ~1.6× fewer steps than Newton–Schulz.
Perturb → Lorenza (TMLR 2026): is a SAM style optimizer for better generalization with a single backprop and low memory.
5/ One object, the spectrum of the gradient, turned out to control memory, speed and generalization.
Read my full post about it (and try the code): ofirlin.substack.com/p/the-s…
All these results are based on a joint work with Yehonathan Refael, Jonathan Svirsky, Boris Shustin, Wasim Huleihel, Guy Smorodinsky, Iftach Arbel and Tom Tirer.
Ofir Lindenbaum retweeted
NYU CDS is hiring its next cohort of Faculty Fellows! 🗽 A chance to do independent research, with a light teaching load and better conditions than a typical postdoc. If you're finishing a PhD in ML/NLP/data science or nearby, apply:
apply.interfolio.com/193498
This is a really cool idea, and a great blog describing it
1/9 Current optimizers are static during training. What if the optimizer could think and react while the network trains?
We let an agent watch the training dynamics and rewrite the optimizer's source code as training goes: a reactive optimizer.
In every language and image setting we tried that has no tuned recipe, it matches or beats the best recipe one can one-shot, and it guards against loss spikes and overfitting by itself. In the near future, optimization decisions may belong to agents, not humans.
📝 xingjianbai.com/reactive-opt…
💻 Try it on your task: github.com/xingjian-bai/opal
🧵
1
21
1,527
Ofir Lindenbaum retweeted
לכבוד ניסוי מספר 50 ב- Wonderlattice, הנה סרטון קצר שעובר על כולם (בעברית) ונותן טעימה קטנה מכל אחד
Ofir Lindenbaum retweeted
I am excited to speak at AI Week 2026 and a little nervous about stepping into the lion's den! Ping me if you are coming, and I would love to meet up!
I'll talk about training and inference of ML models as first-passage processes and show how we can improve them using resetting.
גאה מאוד ב-5 סטודנטים שלי (ועוד אחד שעבדתי איתו) שזכו בפרס על מאמרים שפרסמו בכנס או בכתב עת מוביל!
במסלול האקדמי של סטודנטי מחקר אין הרבה פרגון והכרה בהצלחות, ולכן ממש משמח לראות שהפקולטה החליטה לפרגן להם ככה 💪
אישית אני מרגיש זכות לעבוד עם אנשים מוכשרים כל כך על בעיות שמעניינות אותי (ואותם)
@ubarilan
Ofir Lindenbaum retweeted
*אנשים טובים הכינו מדריך פרקטי שמסביר איך לארגן מבצע הצבעה שכונתי בצורה פשוטה ואפקטיבית.*
הרעיון הוא לרשת את השכונה לפי בניינים וכניסות, ולדאוג שכל השכנים שרוצים להצביע אכן יגיעו לקלפי. זה כולל מיפוי מוקדם, עזרה בהסעות או בנגישות למי שצריך, ועדכון בזמן אמת לאורך יום הבחירות.
הכל מבוסס על כלים חינמיים, בלי תקציב ובלי לפלג את השכונה. העלנו את כל השלבים והטיפים בקובץ המצורף. מוזמנים להסתכל, לשתף הלאה או להתאים את היוזמה לאופי השכונה שלכם.
עוד פרטים: [email protected]
@obenner
27.10 >>> באים להצביע ומנצחים.
*שתפו והפיצו*
This is such a bad analogy:
-Many drugs were discovered by accident
-We don't fully undertand how many drugs work
-Drugs are generally not "aligned" with what we want them to do. They have many effects, some of which we don't even know about
Btw, building an airplane would be a much better analogy :)
This is a slide from Terence Tao's 'Math 2.0' public lecture yesterday evening at Caltech. I am highly confident that, if the cure actually worked, almost any Stage 4 cancer patient would tell you that human comprehension of its mechanism is completely irrelevant to them.
3
21
1,074
Well, time to eat my hat🎩
What incentive do I have to keep volunteering as an AC for @NeurIPSConf if I can’t even register to attend the conference where my paper is presented?
#NeurIPS2026 Registration for Paris was never open
55
2,726
Very impressive! We've tried to solve this task with many tricks (they all failed beyond synthetic data)
What incentive do I have to keep volunteering as an AC for @NeurIPSConf if I can’t even register to attend the conference where my paper is presented?
#NeurIPS2026 Registration for Paris was never open
I checked the abstracts of 6 #ICLR submissions with @pangram. It flagged 5 as 100% AI-generated.
Some were also barely readable. I’m all for using AI to improve writing, but the result still needs careful editing.
Regardless of how accurate @pangram is, an abstract should clearly explain the problem, the contribution, and the findings. This is not the case for most abstracts I read this year @iclr_conf
