Baharan Mirzasoleiman (@baharanm) profile picture

Assistant professor @UCLAComSci. Better ML via better data, Machine learning, Optimization

Los Angeles, CA
Joined July 2018
Photo by Baharan Mirzasoleiman (@baharanm)Photo by Baharan Mirzasoleiman (@baharanm)Photo by Baharan Mirzasoleiman (@baharanm)Photo by Baharan Mirzasoleiman (@baharanm)
Weโ€™re thrilled by the amazing response to our #ICML2024 tutorial on โ€œFoundations of data-efficient learningโ€! Over 1000 attendees joined us. Thank you all! ๐Ÿ™Œ๐ŸŒฑ๐ŸŒฑ๐ŸŒฑ โžก๏ธ Slides: baharanm.github.io/ICML24_tuโ€ฆ โžก๏ธ Recording: will be available on Aug 22 ๐ŸŽŠ๐ŸŽŠ
I'll be giving a 2-hour tutorial on data-efficient learning with my PhD student @sjoshi804 on Monday July 22 at #ICML2024. Join us to learn more about this cool topic! โžก๏ธ We can learn better from better data! โฌ…๏ธ๐Ÿ™Œ๐ŸŒฑ
6
17
215
41,239
Baharan Mirzasoleiman retweeted
Announcing the ๐Ÿ†๐๐ž๐ฌ๐ญ ๐๐š๐ฉ๐ž๐ซ ๐€๐ฐ๐š๐ซ๐๐ฌ๐Ÿ† for DATA-FM workshop @iclr_conf: ๐ถ๐‘Ÿ๐‘Ž๐‘š ๐ฟ๐‘’๐‘ ๐‘  ๐‘ก๐‘œ ๐น๐‘–๐‘ก ๐‘€๐‘œ๐‘Ÿ๐‘’: ๐‘‡๐‘Ÿ๐‘Ž๐‘–๐‘›๐‘–๐‘›๐‘” ๐ท๐‘Ž๐‘ก๐‘Ž ๐‘ƒ๐‘Ÿ๐‘ข๐‘›๐‘–๐‘›๐‘” ๐ผ๐‘š๐‘๐‘Ÿ๐‘œ๐‘ฃ๐‘’๐‘  ๐‘€๐‘’๐‘š๐‘œ๐‘Ÿ๐‘–๐‘ง๐‘Ž๐‘ก๐‘–๐‘œ๐‘› ๐‘œ๐‘“ ๐น๐‘Ž๐‘๐‘ก๐‘  by @jiayuan_ye_ et al. arxiv.org/abs/2604.08519 (morning oral session) ๐Ÿฅˆ๐‡๐จ๐ง๐จ๐ซ๐š๐›๐ฅ๐ž ๐Œ๐ž๐ง๐ญ๐ข๐จ๐ง goes to ๐ด๐‘›๐‘โ„Ž๐‘œ๐‘Ÿ๐‘’๐‘‘ ๐ท๐‘’๐‘๐‘œ๐‘‘๐‘–๐‘›๐‘”: ๐‘ƒ๐‘Ÿ๐‘œ๐‘ฃ๐‘Ž๐‘๐‘™๐‘ฆ ๐‘…๐‘’๐‘‘๐‘ข๐‘๐‘–๐‘›๐‘” ๐ถ๐‘œ๐‘๐‘ฆ๐‘Ÿ๐‘–๐‘”โ„Ž๐‘ก ๐‘…๐‘–๐‘ ๐‘˜ ๐‘“๐‘œ๐‘Ÿ ๐ด๐‘›๐‘ฆ ๐ฟ๐‘Ž๐‘›๐‘”๐‘ข๐‘Ž๐‘”๐‘’ ๐‘€๐‘œ๐‘‘๐‘’๐‘™ by @jcqln_h et al. arxiv.org/abs/2602.07120 (afternoon oral session) Congratulations to all the authors, and please join us at the workshop tomorrow at ๐‘๐จ๐จ๐ฆ ๐Ÿ๐ŸŽ๐Ÿ‘ ๐€+๐ for more discussions around everything data! ๐Ÿ“ Updated schedule at data-fm-iclr2026.github.io/, with invited talks by @sewon__min @baharanm @jcniebles @fredsala! Thank you to our sponsors @SnorkelAI and @datologyai for all your support!
1
9
1
77
7,757
Can weak LLMs supervise strong LLMs to obtain superior performance? ๐Ÿค” Yes!! ๐Ÿคฉ Which weak models are better supervisors? ๐Ÿค” Check out @xue_yihao65785โ€™s awesome #icml2025 paper to know how to identify best weak supervisors without having to collect labels! ๐ŸŽ‰๐ŸŒฑ
๐ŸŽ‰ Our paper โ€œRepresentations Shape Weak-to-Strong Generalizationโ€ is accepted at #ICML2025! We study weak-to-strong generalization (W2SG)โ€”a core problem in superalignmentโ€”and offer new insights into the role of models' internal representations in W2SG. 1/
1
2
25
2,634
๐Ÿšจ Join us at the Workshop on Spurious Correlation & Shortcut Learning (SCSL) at #ICLR2025! @iclr_conf ๐Ÿ—“๏ธ April 28, 2025 ๐Ÿ“ Garnet 214-215, Singapore EXPO ๐ŸŒ More info: scslworkshop.github.io #ICLR2025
16
3
50
9,403
Want to train LLMs with better performance and lower GPU memory requirements on data mixtures? Check out this cool #ICLR2025 paper of @dangnth97 this Friday April 25 at 10am, Poster #265 ๐ŸŽ‰๐ŸŒฑ
๐ŸŽ‰ Achievement unlocked: having papers with all of my labmates and somehow all ended up at ICLR! Iโ€™ll be presenting our work โ€œMini-batch Coresets for Memory-efficient Language Model Training on Data Mixturesโ€ at #ICLR2025 ๐Ÿ‡ธ๐Ÿ‡ฌ Come by and chat! ๐Ÿ‘‹ on Fri, Apr 25 | 10 AM GMT+8
1
19
1,972
Can we pretrain deep models with small synthetic data? Dataset Distillation via Knowledge Distillation is the way to go! Check out @sjoshi804โ€™s #ICLR2025 paper this Saturday April 26 at 9am, Poster #307 ๐ŸŽ‰๐ŸŒฑ
#ICLR2025 Can you pre-train deep models with small, synthetic datasets? ๐Ÿคฏ We introduce the first effective dataset distillation method for self-supervised learning (SSL) โ€” boosting downstream accuracy by up to 13% over baselines. ๐Ÿงช Poster #307, Sat Apr 26, 9am
1
4
37
2,537
Big congrats @YuYang_i on your graduation!! ๐ŸŽ‰๐ŸŽ‰ ๐ŸŽ‰ very nice PhD thesis with great contributions ๐ŸŒฑ Iโ€™m proud of all youโ€™ve done, and I wish you the best! ๐Ÿ’
Sharing a little late update (before itโ€™s no longer news): I wrapped up my PhD at the end of last year and recently joined @OpenAIโ€™s reasoning team ๐Ÿ“โœจ!
56
6,008
Baharan Mirzasoleiman retweeted
(2/2) Not at UCLA but interested in this work? Check arxiv.org/abs/2502.02407. Thanks to our fantastic intern @unregularized (soon to return full-time!) for leading this project, along with collaborators @ynd and Atish. Thanks to my UCLA host @baharanm for the seminar invitation!
2
1
20
2,331
Baharan Mirzasoleiman retweeted
We are delighted that our proposal for the Workshop on โ€œSpurious Correlation and Shortcut Learning: Foundations and Solutionsโ€ has been accepted at @iclr_conf 2025, hosting many brilliant keynote speakers and panelists. Stay tuned: scslworkshop.github.io @SCSLWorkshop 1/
1
16
2
75
18,337
Baharan Mirzasoleiman retweeted
At NeurIPS? Check out the 2nd workshop on Attributing Model Behavior at Scale (ATTRIB)! Meeting Rm 205-207, starting @ 9am - amazing talks from @SurbhiGoel_ @sanmikoyejo @baharanm, Robert Geirhos, and @coallaoh + exciting contributed talks! More info: attrib-workshop.cc/
8
4
39
9,229
Iโ€™ll help presenting our #NeurIPS2024 posters tomorrow (Friday):๐ŸŒฑ 1- Changing the training data distribution to improve in-distribution performance (11@west #7106) w. @dangnth97 2- Data selection for fine-tuning LLMs with superior performance (16:30@west #5401) w. @YUYANG_UCLA
2
5
57
8,655
Baharan Mirzasoleiman retweeted
Attending NeurIPS'24? Please mark your calendar for our special event "SFU@NeurIPS 2024" sites.google.com/view/sfu-atโ€ฆ 9 speakers from both academia & industry! Only a 10-min walk from the convention center! Letโ€™s enjoy exciting talks and open discussions!
2
19
6
105
26,280
Same training and test distribution yields optimal in-distribution performance? @dangnth97 showed in his #NeurIPS2024 paper that this is not true when training with gradient methods!!๐Ÿ˜ฎ๐Ÿ™ƒ Changing the training data distribution yields SOTA!๐ŸŽŠ Check it out Fri Dec 13, 11am, PS#5
1
23
1,235
Smaller high-quality subsets of language data not only improve LLMsโ€™ training efficiency, but also yield considerably better performance! ๐Ÿ™Œ๐ŸŽ‰๐ŸŒฑ @YUYANG_UCLA has a theoretically-rigorous method for this in her #NeurIPS2024 paper! Check it out on Fri, Dec 13, 16:30, #PS 6
1/ I'll be at #NeurIPS2024 presenting our work SmallToLarge (S2L): Data-efficient Fine-tuning of LLMs! ๐Ÿš€ Whatโ€™s S2L? Itโ€™s a scalable data selection method that trains a small proxy model to guide fine-tuning for larger models, reducing costs while preserving performance. ๐Ÿ‘‡
8
49
8,267
Baharan Mirzasoleiman retweeted
๐Ÿ“ข @UCLAComSci is hiring! Open to all CS areas! - Multiple Tenure-track Assistant Professor Positions: recruit.apo.ucla.edu/JPF0979โ€ฆ - Open Rank Teaching Professor Position: recruit.apo.ucla.edu/JPF0980โ€ฆ (We hired 11 Assistant Professors in the past two years ...)
2
26
1
94
17,325
Baharan Mirzasoleiman retweeted
The Adversarial Machine Learning Rising Star Awards deadline is in two weeks! Submit your application and help us promote your work and research vision! @trustworthy_ml @LLMSecurity @ml_safety @safe_paper
๐Ÿšฉ(1/2) Please help forward the Call for the 2024 Adversarial Machine Learning (AdvML) Rising Star Awards! We promote junior researchers in AI safety, robustness, and security. Award events are hosted at AdvML'Frontiers workshop @NeurIPSConf 2024 Info: sites.google.com/view/advml/โ€ฆ
2
12
1,992
Iโ€™ll also present โ€œSafeClipโ€ on behalf of @WenhanYang0315 tomorrow at 1:30pm (poster session 6) #814. See you there! ๐Ÿ™Œ
CLIP is highly sensitive to data poisoning and backdoor attacks. In this #ICML2024 paper, @WenhanYang0315 proposed an interesting way to pretrain CLIP robust to such attacks without compromising the performance! ๐ŸŒฑ๐ŸŒฑ ๐Ÿ”—arxiv.org/pdf/2310.05862 Thu, July 25, Poster session 6, #814
2
18
1,678
Iโ€™ll present โ€œMixProโ€ on behalf of @xue_yihao65785 tomorrow at 11:30 (poster session 5) poster #800. Come check it out ๐Ÿ™Œ
ML models are sensitive to distribution shift. Can we adapt a model with only a few examples from the target domain? In this #ICML2024 paper, @xue_yihao65785 proposes an effective way, with nice theoretical analysis๐ŸŒฑ ๐Ÿ”—arxiv.org/pdf/2305.14521 Thu, July 25, Poster session 5, #800
1
18
1,780