AI Paper Atlas

Research map / Optimization / Implicit bias and generalization

Understanding the Dynamics of Gradient Flow in Overparameterized Linear models

Salma Tarmoun, Guilherme Franca, Benjamin D Haeffele, Rene Vidal

ICML 2021 · Optimization › Implicit bias and generalization · approach: Gradient flow and mean-field dynamics

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Initialize overparameterized factorization with larger conserved imbalance ||Q0|| (balanced 0.02 -> imbalanced 196.55)two-layer linear model (asymmetric factorization, gradient descent) · random Gaussian Y, m=5, n=10, k=50, syntheticreconstruction error convergence speedImproved
Increase imbalance of initialization (||Q0|| from 0.02 to 1.83) (ablation)two-layer linear model · random Gaussian Y, m=5, n=10, k=50, syntheticconvergence speed of singular valuesImproved
Use balanced (near-zero conserved quantity) initialization instead of imbalanced (ablation)two-layer linear model · random Gaussian Y, syntheticorder of singular value learningNo clear effect
Add imbalance-driven conservation law (nonzero Q0) to nonlinear two-layer sigmoid networknonlinear two-layer network · synthetic data, n=1000, d=10, width 100training lossNo clear effect
Replace the 1/8 weight of the regularizer in explicit-regularization dynamics by a different constant alpharegularized factorization dynamics · theoryequivalence to conservation-law dynamicsHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy