A mean-field theory for multi-component online ICA in high dimensions predicts decoupled and competition phases, explicit learnability boundaries, and a staircase effect in the number of recoverable components as a function of learning rate.
Natural gradient works efficiently in learning.Neural Computation, 10(2):251– 276
2 Pith papers cite this work. Polarity classification is still indexing.
2
Pith papers citing it
citation-role summary
background 1
citation-polarity summary
years
2026 2verdicts
UNVERDICTED 2roles
background 1polarities
background 1representative citing papers
Ember is a memory-efficient optimizer for token embeddings that exploits distinct gradient geometry and models token trajectories as 1D rays.
citing papers explorer
-
Learnability and Competition in High-Dimensional Multi-Component ICA
A mean-field theory for multi-component online ICA in high dimensions predicts decoupled and competition phases, explicit learnability boundaries, and a staircase effect in the number of recoverable components as a function of learning rate.
-
Token Geometry
Ember is a memory-efficient optimizer for token embeddings that exploits distinct gradient geometry and models token trajectories as 1D rays.