AI Paper Atlas

Research map / Architectures and efficiency / CNN design

Decoupled Greedy Learning of CNNs

Eugene Belilovsky, Michael Eickenberg, Edouard Oyallon

ICML 2020 · Architectures and efficiency › CNN design · approach: Theory of convolutional inductive bias · 32 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
Replace CNN-aux with MLP-aux (ablation)DGL CNN · CIFAR-10accuracy: 92.2 -> 90.6Hurt
Replace CNN-aux with MLP-aux (ablation)DGL CNN · CIFAR-10relative FLOPS: 200% -> 0.7%Improved
Use staged spatial reduction in the auxiliary network instead of CNN-aux (ablation)DGL CNN · CIFAR-10accuracy: 92.2 -> 91.2Hurt
Use staged spatial reduction in the auxiliary network instead of CNN-aux (ablation)DGL CNN · CIFAR-10relative FLOPS: 200% -> 4.0%Improved
Optimize greedy layers in parallel rather than sequentially (ablation)DGL · CIFAR-10, 6-layer CNNaccuracyNo clear effect
Use DGL rather than DDG for module trainingResNet-110 · CIFAR-10 (K=2)accuracy: 93.41 vs 93.5±0.1No clear effect
Use asynchronous scheduling with a slowed layer up to approximately 1.2× (ablation)Async DGL · CIFAR-10, 5-layer CNNaccuracy: approximately 1.2×No clear effect
Use smaller replay buffers (ablation)Async DGL · CIFAR-10, 5-layer CNNaccuracyHurt
Use DGL rather than end-to-end backpropagationCNNs · CIFAR-10, CNN experiment comparing DNI methodstest accuracyImproved

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