Self-Supervised Learning for CIFAR-10 Image Classification

The classification of images is one of the most important tasks of the machine vision processes, and deep learning algorithms like CNNs are used to recognize given patterns within an image. Unfortunately, it is also expensive and time-consuming to collect labeled data. In this paper, we make an attempt to address this issue by presenting a new method based on Self-Supervised Learning (SSL). Models can be trained on tasks that do not require a label and thus decrease the consumption of training data. The first method proposed is ‘multi-task rotation prediction’, where the model rotates different images and the second method is a jigsaw puzzle task rearranging shuffled patches of images and correcting the color channels. This method details a new approach to image classification using unlabeled data, which can have far reaching consequences in domains that do not have easily available labeled datasets.