Hiroki Naganuma

Overall

Work in Progress

Until Friday, I was staying at Google DeepMind Japan (as work from anywhere) and join end year party of Masason Foundation, so my participation in NeurIPS was limited to the two days of workshops on Saturday and Sunday. Therefore, I would like to focus mainly on these two days in my summary.

Networking + Seeing Friends

Day1

• Toyotaro Suzumura / UTokyo
• Shinji Ito / RIKEN
• Masatoshi Uehara / University of Wisconsin-Madison
• Hao Chen / CMU
• Bai Cong / Science Tokyo
• Ryosuke Yamaki / URitsumeikan
• Kotaro Yoshida / Science Tokyo

• Sabyasachi Sahoo / ULaval+Mila

• Mila
  • Reyhane Askari Hemmat Meta
  • Mohammad Pezeshki / Meta

Day2

• Shikai Qiu / NYU, GDM (work w. Atish)
• Jack Min Ong / Prime Intellect (Author of OpenDiLoCo)
  • We talked about my work at MSR (which is related to OpenDiLoCo)

• Kazuto Fukuchi / UTsukuba

• Mila
  • Ryan D’Orazio / Mila
  • Mehrnaz Mofakhami / Mila
  • Charles Guille-Escuret / Mila

FITML Workshop

• url

I will present a work "Mastering Task Arithmetic: τJp as a Key Indicator for Weight Disentanglement" at the #NeurIPS2024 workshop FITML on 14th Dec.

We mitigate the interference between task vectors and coefficient tuning costs through the "τJp" regularization. pic.twitter.com/weaj9Ba75v

— Kotaro Yoshida @NeurIPS2024 (@katoro13___) December 12, 2024

Papers

WIP

OPT Workshop

• url

I will present our work on the efficient distributed training algorithm at the optimization workshop. Join us during our poster sessions from 15:00-16:00.#NeurIPS2024 pic.twitter.com/D4nA8hWX4u

— Hiroki Naganuma (@_Hiroki11x) December 15, 2024

Papers

• Pseudo-Asynchronous Local SGD: Robust and Efficient Data-Parallel Training
• Addax: Utilizing Zeroth-Order Gradients to Improve Memory Efficiency and Performance of SGD for Fine-Tuning Language Models
• Scaling Collapse Reveals Universal Dynamics in Compute-Optimally Trained Neural Networks
• Second-Order Forward-Mode Automatic Differentiation for Optimization
• Don’t Be So Positive: Negative Step Sizes in Second-Order Methods
• Scaling Collapse Reveals Universal Dynamics in Compute-Optimally Trained Neural Networks

Spotlight and Keynotes

• SOAP: Improving and Stabilizing Shampoo using Adam

  • Benchmark
    • CIFAR-5M, ImageNet
  • Memo
    • SOAP has better critical batchsize
    • In language modling task, decoupling the eigenvalue and eigenbasis improves the performance

• μLO: Compute-Efficient Meta-Generalization of Learned Optimizers

  • related work: VeLO: Training Versatile Learned Optimizers by Scaling Up
  • github

• Catapults in SGD: spikes in the training loss and their impact on generalization through feature learning

  • Mikhail (Misha) Belkin
  • small batch -> more catapults -> AGOP aignment -> better generalization
  • supported by only empirical side

Topics related to schedulle free, learning rate schedule

WIP

Acknowlegements

I want to thank Microsoft Research (YVR->SFO) and Masason Foundation (SFO->NRT->YVR) for supporting my participation in the NeurIPS.