AI Paper Atlas

Research map / Optimization / Implicit bias and generalization

Phase diagram of early training dynamics in deep neural networks: effect of the learning rate, depth, and width

Dayal Singh Kalra, Maissam Barkeshli

NeurIPS 2023 · Optimization › Implicit bias and generalization · approach: Edge-of-stability 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
Increase depth d and 1/widthSGD on FCN/CNN/ResNet, MSE · early training phase diagramc_loss, c_sharp: c_loss 2.82 -> 36.75Improved
Increase depth and 1/width opens a sharpness-reduction phase between c_crit and c_lossSGD on DNNs · early trainingsharpnessImproved
Center the network output at initialization (f - f(theta0)) (ablation)FCN d=8 CIFAR-10 MSE · phase diagramsharpness reduction phase / c_lossImproved
Set last layer weights to zero at initialization (ablation)FCN d=8 CIFAR-10 MSE · phase diagramsharpness reduction phase / c_lossImproved
Scale the output by a fixed factor alpha (ablation)FCN · early trainingc_loss growth with d and 1/wNo clear effect
Use cross-entropy instead of MSESGD · early trainingc_loss and c_crit: c_crit 2 -> 4 >= c_crit >= 2Mixed
Reduce batch size from 512 to 32 (ablation)SGD · early training phase diagramphase diagramNo clear effect
Add bias (ablation)SGD networks · early trainingphase diagramNo clear effect
Increase width of uv model from w=2 to w=512 (ablation)uv model · MSE, y=0c_loss: 3.74 -> 2.1Improved
Increase learning rate constant c above c_crit7-layer CNN · CIFAR-10 sharpness at c tau=200normalized sharpnessImproved
Increase learning rate between c_loss and c_barrierCNNs · linear interpolation of lossloss barrierNo clear effect

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