Figureure 10 · Summary of recognition performance with $\rho ( z ) = R e L U .$Figure 10. Summary of recognition performance with $\rho ( z ) = R e L U .$ . Testing is always performed on CIFAR10. (left) training is performed on different datasets using crop translation as augmentation. (right) training is done on CIFAR10 using different augmentations. The yellow horizontal line corresponds to random weights. The red horizontal line is partial whitening in PCA space (Thiry, 2021). Even with such a simple architecture, CL learns representations that lead to improved recognition accuracy but most of that improvement is due to partial whitening. The usefulness of training on random images for recognizing real images is best for noise images that approximatel match the power spectrum of CIFAR10.这张图概括 A Theory of Contrastive Learning 的整体方法流程。阅读时先看模块之间传递的训练信号，再看作者如何把目标拆成可优化的子问题。Figureure 17 · Summary of our experimental results with different augmentations and qFigure 17. Summary of our experimental results with different augmentations and quadratic and ReLU nonlinearities. This figure extends the content of fig. 8 by adding the sensitivities in addition to the filters. Results are after 2000 training epochs.这张图/表用于判断 A Theory of Contrastive Learning 的实验收益来自哪里。重点看替换、消融或跨模型设置下趋势是否一致，而不是只看单个最高分。
