Label-Augmented Dataset Distillation (LADD)
收藏资源简介:
Label-Augmented Dataset Distillation (LADD) 是一个专注于标签增强的数据集蒸馏框架,旨在通过标签增强提升数据集蒸馏的效果。该数据集通过子采样合成图像并生成密集标签来捕捉丰富的语义信息,仅增加2.5%的存储空间,显著提升性能。LADD的创建过程包括蒸馏和部署两个阶段,通过合并全局视图图像和局部视图图像及其对应的密集标签,提供多样化的学习信号。该数据集主要应用于提升模型训练效率和性能,特别是在跨架构的鲁棒性方面表现出色。
Label-Augmented Dataset Distillation (LADD) is a label-augmented dataset distillation framework designed to improve the performance of dataset distillation via label augmentation techniques. It captures rich semantic information by subsampling synthetic images and generating dense labels, incurring only a 2.5% increase in storage overhead while significantly boosting model training efficacy and overall performance. The development of LADD comprises two stages: distillation and deployment. It delivers diverse learning signals by integrating global-view images, local-view images and their corresponding dense labels. This dataset is primarily utilized to enhance the efficiency and performance of model training, and performs exceptionally well in terms of cross-architecture robustness.

- 1Label-Augmented Dataset Distillation延世大学, KAIST AI · 2024年



