Nicolas Loizou
@NicLoizou
Followers
1K
Following
742
Statuses
168
Assistant Professor @JohnsHopkins. Optimization and Machine Learning.
Joined June 2020
Not to be missed! Check out Sayantan's recent work on a Scale-Invariant Version of Adagrad! He will present this at NeurIPS this week: Wed 11 Dec 11 a.m.- 2 p.m. PST He’s also on the lookout for exciting opportunities starting next summer!
(1/4) Excited to share our NeurIPS 2024 paper: Remove that Square Root: A New Efficient Scale-Invariant Version of AdaGrad 🚀 🤝work with @dr_nazya @NicLoizou @sam_hrvth @TakacMartin @ed_gorbunov 📍Poster: West Ballroom A-D #6103 📅 Wed 11 Dec 11 a.m.- 2 p.m. PST
0
1
10
RT @ErnestRyu: @shuvo_das_gupta (Columbia) and I (UCLA) are starting an optimization seminar series! Our first speaker, @aaron_defazio (Met…
0
10
0
Applications are now open for the Data Science and AI Institute Postdoctoral Fellowship Program @HopkinsDSAI for the 2025-2026 academic year. Excellent environment, great benefits, and many researchers to collaborate with:
0
6
18
Are you looking for a great place to pursue your PhD in Canada? Check out @Mila_Quebec! It's a fantastic environment to advance research at the intersection of optimization and machine learning alongside amazing colleagues. @bouzoukipunks, @SimonLacosteJ, @gauthier_gidel
Mila's annual supervision request process opens on October 15 to receive MSc and PhD applications for Fall 2025 admission! Join our community! More information here
0
0
7
RT @BachFrancis: How fast is gradient descent, *for real*? Some (partial) answers in this new blog post on scaling laws for optimization. h…
0
85
0
#neurips2024 In one of the most unexpected decisions, our paper, with scores 6, 6, 6, and 7 (average rating: 6.25, confidence: 3), was rejected with no reasonable justification from the AC. @jmtomczak @ulrichpaquet @DaniCMBelg ACs should be held accountable for such decisions.
#neurips2024 An adversarial AC rejected our paper with avg score of 6.3 with ALL Reviewers accepted the paper because "the reviewers were not responding on a clear paper during discussion" WTH is going on? @jmtomczak @ulrichpaquet @DaniCMBelg Any accountability for such an AC?
6
6
54
RT @JohnsHopkins: Johns Hopkins University achieves its highest-ever position in the U.S. News & World Report rankings coming in at No. 6 f…
0
71
0
RT @deepmath1: Join us in Philadelphia at the 8th iteration of the conference on mathematical theory of deep neural networks @deepmath1, su…
0
7
0
@bremen79 @gautamcgoel @MatharyCharles For this, check out the upper conditions in the paper: I believe your exact condition might be related to one or two conditions there. Remark: Your condition could also be related to the Polyak steps! ;)
1
0
3
I am looking forward to further extensions of Schedule-Free Algorithms in the future! Many Congrats to @aaron_defazio @konstmish @alicey_ang! Well-deserved recognition!
Schedule-Free Wins AlgoPerf Self-Tuning Track 🎉 I'm pleased to announce that Schedule-Free AdamW set a new SOTA for self-tuning training algorithms, besting AdamW and all other submissions by 8% overall. Try it out:
0
2
16
On my way to #ISMP2024! Tomorrow, I’ll be presenting our latest findings on Stochastic Extragradient Methods at On Thursday, don’t miss our sessions on: "Adaptive, Line-search, and Parameter-Free Stochastic Optimization".
1
2
24
@orvieto_antonio @konstmish Is NGN provably at most twice the SPSmax value? If you fix the problem, then will 1.5*SPSmax have a similar performance to this? In other words, is the beneficial performance an outcome of the slightly larger step-size selection?
2
0
1
I have been waiting for this to drop since Antonio's excellent PhD thesis! Looking forward to exploring the connections with Polyak steps and extensions to coordinate-wise and momentum variants!
NGN introduces a new paradigm for optimization in deep learning: cheap curvature estimation with state-of-the-art theory. With Lin Xiao, we took one year to write this - it was just too good to rush. A lot still has to be done, this is just the beginning.
1
0
11
1
0
0
All of our convergence guarantees are true generalizations of the theoretical analysis of SGD (no momentum) using SPSmax, DecSPS, and AdaSPS. Finally, we include numerical experiments on various problems, demonstrating the effectiveness and practicality of the new update rules.
0
0
3