Dataset Distillation
收藏资源简介:
数据集蒸馏是一种创新的数据处理技术,由麻省理工学院计算机科学与人工智能实验室开发。该技术旨在将大型训练数据集的知识压缩到一个小型合成数据集中,例如将60,000个MNIST训练图像压缩为仅10个合成图像(每类一个)。这些合成图像虽不源自原始数据分布,但用于训练时能接近原始数据集的性能。数据集蒸馏技术适用于多种初始化设置和学习目标,已在多个公开数据集上展示了其优越性,主要用于解决数据压缩和模型训练效率问题。
Dataset Distillation is an innovative data processing technique developed by the Computer Science and Artificial Intelligence Laboratory (CSAIL) of the Massachusetts Institute of Technology (MIT). This technique aims to compress the knowledge embedded in large-scale training datasets into a small synthetic dataset. For instance, it can compress 60,000 MNIST training images into merely 10 synthetic images (one per class). Although these synthetic images are not sampled from the original data distribution, they can achieve performance comparable to that of the original dataset when used for model training. Dataset Distillation is compatible with various initialization settings and learning objectives, and its superiority has been demonstrated on multiple public datasets. It is primarily utilized to address issues of data compression and model training efficiency.




