2020/10/09 by Oren Nuriel, Nuriel, Oren, Sagie Benaim +3 · 2 citations
Medicine · Computer Science · #COVID-19 diagnosis using AI #Domain Adaptation and Few-Shot Learning #Image Retrieval and Classification Techniques
paper · pdf · doi:10.48550/arxiv.2010.05785
Recent work has shown that convolutional neural network classifiers overly\nrely on texture at the expense of shape cues. We make a similar but different\ndistinction between shape and local image cues, on the one hand, and global\nimage statistics, on the other. Our method, called Permuted Adaptive Instance\nNormalization (pAdaIN), reduces the representation of global statistics in the\nhidden layers of image classifiers. pAdaIN samples a random permutation \π\nthat rearranges the samples in a given batch. Adaptive Instance Normalization\n(AdaIN) is then applied between the activations of each (non-permuted) sample\ni and the corresponding activations of the sample \π(i), thus swapping\nstatistics between the samples of the batch. Since the global image statistics\nare distorted, this swapping procedure causes the network to rely on cues, such\nas shape or texture. By choosing the random permutation with probability p\nand the identity permutation otherwise, one can control the effect's strength.\n With the correct choice of p, fixed apriori for all experiments and\nselected without considering test data, our method consistently outperforms\nbaselines in multiple settings. In image classification, our method improves on\nboth CIFAR100 and ImageNet using multiple architectures. In the setting of\nrobustness, our method improves on both ImageNet-C and Cifar-100-C for multiple\narchitectures. In the setting of domain adaptation and domain generalization,\nour method achieves state of the art results on the transfer learning task from\nGTAV to Cityscapes and on the PACS benchmark.\n