Distilled Datasets

Distilled Datasets是通过应用一种新颖的基于核的元学习框架创建的，用于数据集蒸馏，特别是在使用无限宽卷积神经网络时。这些数据集在多个图像分类任务上实现了最先进的结果，例如在CIFAR10图像分类任务上，仅使用10个数据点（原始数据集的0.02%）就能达到超过65%的测试准确率。这些数据集不仅在大小上显著减小，而且在性能上保持高效，适用于核岭回归和神经网络训练，为训练效率和有用特征提取提供了价值。

Distilled Datasets are constructed via a novel kernel-based meta-learning framework tailored for dataset distillation, especially when leveraging infinitely wide convolutional neural networks. These datasets achieve state-of-the-art performance across multiple image classification tasks. For instance, on the CIFAR-10 image classification task, they attain over 65% test accuracy with merely 10 data points (0.02% of the original dataset). Not only do these datasets achieve significant size reduction, but they also maintain high performance efficiency, being applicable to both kernel ridge regression and neural network training, providing value for training efficiency and effective feature extraction.