Percolation Dynamics in Optimization: Variance Cascades in SGD, Adam and AdamW

Sai Niranjan Ramachandran · hf · 2026-09-04

SGD dynamics are modeled as a percolation process where architectural symmetries cause subnetworks to merge in discrete blocks, producing variance spikes resembling phase transitions; similar trapping behavior under heavy-tailed noise extends to Adam and AdamW.

Original post →

More from Research

Research channel →