Skip to content

Muon is bad at memorization? #13

Description

@KindXiaoming

With a memorization dataset, it's hard to find a good learning rate for Muon, either too small (progress is slow), or too large (large oscillations). By contrast, a SOAP or Adam optimizer (leave for readers to try) can nicely work for memorization dataset.

Muon (lr = 0.01)

Image

Muon (lr = 0.1)

Image

How to reproduce

  1. open comfyResearch: https://online.comfy-research.com/
  2. Use Lottery mode and Type in '142859' in the input field
  3. draw the random graph and run

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    open questionInteresting phenomenon call for explanations

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions