Improving image classification robustness using self?supervision

Self-supervised learning allows training of neural networks without immense, high-quality or labeled datasets. We demonstrate that self-supervision furthermore improves robustness models using small, imbalanced incomplete datasets which pose severe difficulties to supervised models. For small datasets, the accuracy our approach is up 12.5% higher MNIST and 15.2% Fashion-MNIST compared random initialization. Moreover, influences way itself, means in case strongly it can be prevented classes are not insufficiently learned. Even if input data corrupted large image regions missing from trainingset, significantly classification (up 7.3% for 2.2% Fashion-MNIST). In addition, we analyze combinations manipulations seek generate a better understanding how pretext downstream related. This only important ensure optimal pretraining, but also with unlabeled order find an appropriate evaluation measure. As such, make contribution realistic making machine accessible application areas require expensive difficult collection.