2022.3.7 论文速览
接下来的论文解读可能穿插更多的英文表述,懒,多多谅解。
文章目录
CAFE: Learning to Condense Dataset by Aligning Features(CVPR2022)
Method
本文主要目的是寻找一个替代数据集来减少模型的计算开销,具体做法是构建了一个合成数据集,构建方法如下:
Layer-wise feature alignment: minimize the difference of layer-wise feature maps of real and synthetic images using Mean Square Error (MSE).
To enable learning discriminative synthetic images, we use the feature centers of synthetic images of each class to classify the real images by computing their inner-product and cross-entropy loss.
Experiments
核心子集精度
可视化结果
HCSC: Hierarchical Contrastive Selective Coding(CVPR2022)
Method
现有的对比学习方法没有考虑到图片的分层特性
Experiments
框架结构:有一说一,这个分层的先验知识要求会不会太强了
部分实验结果: