AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Feature-Learning Networks Are Consistent Across Widths At Realistic Scales

Nikhil Vyas, Alexander Atanasov, Blake Bordelon, Depen Morwani, Sabarish Sainathan, Cengiz Pehlevan

NeurIPS 2023 · Architectures and efficiency › RNNs and normalization · approach: Theory of depth and width · 3 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 width in μP ResNet-18ResNet-18 · CIFAR-5mtraining crossentropy consistencyImproved
Increase width in μP ResNet-18ResNet-18 · ImageNettraining crossentropy consistencyImproved
Increase width in μP transformerTransformer · Wikitext-103training crossentropy consistencyImproved
Increase width in μP modelsμP networks · ImageNet and Wikitext-103training crossentropyImproved
Ensemble narrow networks across initializationsμP networks · Online vision and language taskstraining crossentropyMixed
Train for multiple epochs instead of online trainingμP networks · CIFAR-200ktraining error consistency across widthsHurt
Train ensembles on noisy labels for multiple epochsμP ResNets · Noisy CIFAR-50ktest errorMixed
Use μP instead of NTK parameterization (ablation)CNN · CIFAR-10width consistency of test error and kernel alignmentImproved
Use μP instead of standard parameterization (ablation)MLP · Quadratic polynomial tasktraining-loss consistency across widthsImproved
Use μP instead of standard parameterization (ablation)CNN · CIFAR-5mlarge-learning-rate loss-curve consistencyImproved
Use μP instead of standard parameterization (ablation)Transformer · Wikitext-103training-loss consistency across widthsImproved
Use finite-width kernels, which shift more task weight into slower eigenmodesEnsembled eNTK and trained ResNet after-kernel · Polynomial regression and CIFAR-5mkernel-flow lossHurt

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