Adam breaks rotational invariance of factored-model loss

In factored models W = UV^T, the loss is invariant to rotations (U,V)→(UQ,VQ); gradient descent respects that, but Adam's per-coordinate second moment depends on the basis and does not.
1 source
Daily brief
Get tomorrow's AI brief in your inbox
More stories today
- GOP panics over Big Tech ties as Trump shifts on AI regulation
- Ethan Mollick: Claude's skill creator beats ChatGPT for reusable skills
- Aident Loadout gives agents 27,000+ tools and logs every action
- Etched gains sizable fan base for AI inferencing computers
- Corbell generates technical specs from repository knowledge graphs