Researchers have modeled the dynamics of Stochastic Gradient Descent (SGD) as a percolation process, revealing how architectural symmetries cause subnetworks to merge in discrete blocks. These transitions manifest as variance spikes in a macroscopic order parameter, similar to physical phase transitions. The study demonstrates that this trapping mechanism and its associated scaling cascade also apply to Adam and AdamW optimizers under a heavy-tailed noise model. AI
IMPACT Provides a new theoretical framework for understanding optimization dynamics in deep learning.
RANK_REASON The cluster contains a single academic paper published on arXiv. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →