Milking CowMask for Semi-supervised Image Classification

Abstract

Consistency regularization is a technique for semi-supervised learning that has recently been shown to yield strong results for classification with few labeled data. The method works by perturbing input data using augmentation or adversarial examples, and encouraging the learned model to be robust to these perturbations on unlabeled data. Here, we evaluate the use of a recently proposed augmentation method, called CowMask (French et al., 2019), for this purpose. Using CowMask as the augmentation method in semi-supervised consistency regularization, we establish a new state-of-the-art result on Imagenet with 10% labeled data, with a top-5 error of 8.76% and top-1 error of 26.06%. Moreover, we do so with a method that is much simpler than alternative methods. We further investigate the behavior of CowMask for semi-supervised learning by running many smaller scale experiments on the small image benchmarks SVHN, CIFAR-10 and CIFAR-100, where we achieve results competitive with the state of the art, and where we find evidence that the CowMask perturbation is widely applicable. We open source our code at https://github.com/ google-research/google-research/ tree/master/milking_cowmask.

MixMo: Mixing Multiple Inputs for Multiple Outputs via Deep Subnetworks

“…Such non-linear binary masking improves generalization [82,84] by increasing dataset: it creates new images with usually disjoint patches [33]. [4,20,52] seek more diverse transformations via arbitrarily shaped masks: proposals range from cow-spotted masks [24] to masks with irregular edges [33]. As masking of discriminative regions may cause label misallocation [30], [45,46,88,92] try to alleviate this issue with costly saliency heatmaps [76].…”

“…2 compares performance for several mixing blocks [20,33,82,101]. No matter the shape (illustrated in Appendix 6.7), binary masks perform better than linear mixing: the cow-spotted mask (84.17%, 0.561) [23,24] notably performs well. The basic CutMix patching (84.38%, 0.563) is nevertheless more accurate and was our main focus.…”

“…The binary mask is made of the top-λ percentile of pixels from a low-pass filtered 2D map G drawn from an isotropic Gaussian distribution. CowMix [23,24] selects a cow-spotted set of regions, and is somehow similar to FMix with a Gaussian filtered 2D map G. CutMix [101] was inspired by CutOut [16]. Formally, we sample a square with edges of length R √ λ, where R is the length of an image edge.…”

“…The seminal work Mixup [104] linearly interpolates pixels while Manifold Mixup [91] interpolates latent features in the network. Binary masking MSDAs [24,33,46] such as CutMix [101] have since diversified mixed samples by pasting patches from one image onto another in place of interpolation. Aggregating diverse predictions from several neural networks significantly improves generalization [17,32,50,98], notably uncertainty estimation [3,31,69].…”

ClassMix: Segmentation-Based Data Augmentation for Semi-Supervised Learning

“…First, we examine the effect of using different mixed sample data augmentations. Apart from ClassMix we try CutMix [8], as used for semi-supervised semantic segmentation in [6], and CowMix, introduced by French et al [32]. We note that CowMix is very similar to the concurrent FMix, introduced by Harris et al [33].…”