Discriminative Unsupervised Feature Learning with Convolutional Neural Networks

摘要: Current methods for training convolutional neural networks depend on large amounts of labeled samples supervised training. In this paper we present an approach a network using only unlabeled data. We train the to discriminate between set surrogate classes. Each class is formed by applying variety transformations randomly sampled 'seed' image patch. find that simple feature learning algorithm surprisingly successful when applied visual object recognition. The representation learned our achieves classification results matching or outperforming current state-of-the-art unsupervised several popular datasets (STL-10, CIFAR-10, Caltech-101).