A new research paper published on arXiv analyzes the high-dimensional training dynamics of shallow neural networks with quadratic activation functions. The study focuses on the extensive-width regime, where network width scales with input dimension, and uses dynamical mean-field theory to characterize gradient flow. The findings offer quantitative insights into how overparameterization affects learning and generalization, revealing a double descent phenomenon in the presence of label noise and providing an exact expression for the perfect recovery threshold under l2-regularization. AI
IMPACT Provides theoretical understanding of overparameterization effects on neural network learning and generalization.
RANK_REASON The cluster contains a research paper published on arXiv detailing theoretical analysis of neural network dynamics. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
