[cnn]| translation invariance

Input to CNN Why doesn't the image in have translation invariance

^Why do deep convolutional networks generalize so poorly to small image transformations? https://arxiv.org/pdf/1805.12177.pdf
^Making Convolutional Networks Shift-Invariant Again https://arxiv.org/pdf/1904.11486v1.pdf
As demonstrated above CNN Translation invariance cannot be guaranteed in architecture , And the observation shows that , It can only learn some invariance from a large amount of data , The paper believes that there is a certain photographer bias in these general data sets , So that the neural network does not need to learn the formal translation invariance . From a macro perspective , As long as it is not pixel level encoding , There are no two identical images in the world , So neural network can not learn strict translation invariance , There is no need to learn .