AI Paper Atlas

Research map / Architectures and efficiency / CNN design

Dataset Distillation with Infinitely Wide Convolutional Networks

Timothy Nguyen, Roman Novak, Lechao Xiao, Jaehoon Lee

NeurIPS 2021 · Architectures and efficiency › CNN design · approach: Large and dynamic kernel design · 21 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
Use a convolutional NTK rather than the strongest listed prior distillation baselineDataset distillation · CIFAR-10, 1 image/classtest accuracy: 40.5±0.4 -> 64.7±0.2Improved
Use regularized ZCA preprocessing (ablation)KIP ConvNet · CIFAR-10, 10 images/classKIP accuracy: 61.0±0.2 -> 75.6±0.2Improved
Optimize support labels instead of keeping them fixed (ablation)KIP ConvNet · CIFAR-10, 10 images/class, ZCA and no augmentationKIP accuracy: 67.6±0.3 -> 75.6±0.2Improved
Add target-data augmentation (ablation)KIP ConvNet · CIFAR-10, 10 images/class, ZCA and fixed labelsKIP accuracy: 67.6±0.3 -> 66.5±0.3Hurt
Add target-data augmentation (ablation)KIP ConvNet · Fashion-MNIST, 10 images/classKIP accuracy: 91.0±0.1 -> 88.8±0.1Hurt
Train with a family of kernels rather than a single kernelSampling KIP · CIFAR-10, 100 imagestest performance across kernelsImproved
Transfer KIP-distilled data from KRR to finite-width neural network trainingKIP ConvNet · CIFAR-10, 10 images/classtest accuracy: 53.1±0.5 -> 62.7±0.3Improved
Use label-solved data for finite-width NN training instead of KIP dataCIFAR-10 distilled dataset · CIFAR-10, 10 images/classtest accuracy: 62.7±0.3 -> 49.3±0.1Hurt
Use label-solved data for finite-width NN training instead of KIP dataSVHN distilled dataset · SVHN, 10 images/classtest accuracy: 75.0±0.1 -> 53.2±0.3Hurt

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