AI Paper Atlas

Research map / Architectures and efficiency / CNN design

The Singular Values of Convolutional Layers

Hanie Sedghi, Vineet Gupta, Philip M. Long

ICLR 2019 · Architectures and efficiency › CNN design · approach: Theory of convolutional inductive bias · 127 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
Add exact operator-norm projection with clip value 0.5 every 100 steps32-layer ResNet · CIFAR-10test error: 6.2% -> 5.3%Improved
Use clip value 1.0 rather than 0.5 (ablation)Operator-norm projection on 32-layer ResNet · CIFAR-10test error: 5.3% -> 5.5%Hurt
Use clip value 0.1 rather than 0.5 (ablation)Operator-norm projection on 32-layer ResNet · CIFAR-10test error: 5.3% -> 6.7%Hurt
Use clip values 2.5 or 3.0 rather than 0.5 (ablation)Operator-norm projection on 32-layer ResNet · CIFAR-10test errorNo clear effect
Remove batch normalization while testing operator-norm regularization (ablation)32-layer ResNet · CIFAR-10test errorMixed
Replace exact convolution-operator clipping with reshaped-K clipping every 100 steps (ablation)32-layer ResNet · CIFAR-10test accuracyNo clear effect
Clip reshaped-K singular values every step instead of every 100 steps (ablation)Reshaped-K clipping · CIFAR-10test accuracyHurt
Use exact operator-norm clipping instead of reshaped-K clipping32-layer ResNet · CIFAR-10, 10000 batchestraining time: 14490 seconds -> 11004 secondsImproved

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