AI Paper Atlas

Research map / Optimization / Implicit bias and generalization

Surprising Instabilities in Training Deep Networks and a Theoretical Analysis

Yuxin Sun, Dong Lao, Ganesh Sundaramoorthi, Anthony Yezzi

NeurIPS 2022 · Optimization › Implicit bias and generalization · approach: Gradient flow and mean-field dynamics · 1 citations (OpenAlex)

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
perturb last floating-point bit of SGD update (k=3..11 vs k=1)ResNet-56 SGD · CIFAR-10test accuracy STD: 0.16Hurt
replace ReLU with Swish activation (ablation)perturbed SGD, ResNet-56 · CIFAR-10test accuracy variance: 93.72 ± 0.15No clear effect
use different network and dataset (VGG16, Fashion-MNIST) (ablation)perturbed SGD · Fashion-MNISTtest accuracy varianceNo clear effect
decrease fixed learning rate below a minimum (ablation)perturbed SGD, ResNet-56 · CIFAR-10 weight divergencerelative L1 weight differenceImproved
learning rate dt < 0.03 in non-linear PDE (ablation)one-layer CNN PDE · single imageerror amplificationImproved
learning rate 0.03 < dt <= 0.1 in non-linear PDE (ablation)one-layer CNN PDE · single imageerror amplificationHurt
learning rate dt > 0.1 in non-linear PDE (ablation)one-layer CNN PDE · single imagelossHurt
weight decay inside (alpha_min, alpha_max) in linear PDE (ablation)linearized gradient descent PDE · checkerboard inputloss convergenceImproved
finite fixed-support kernel window (16x16 to 64x64) vs infinite support (ablation)linearized PDE · stability bounds on alphaalpha_max: 2e8No clear effect
increase kernel size from 16x16 to 64x64 (ablation)windowed linear PDE · stability bounds on alphaempirical alpha_min: 3.8e6 -> 3.8e7Mixed
vary perturbation constant k (1,3,5,7,9,11) (ablation)perturbed SGD · CIFAR-10final test accuracy: 92.99 to 93.87Mixed

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