Semi-supervised learning via triplet network based active learning

Abstract

Deep learning systems require a large amount of labelled training dataset. However large amount of labelled data is not available in many cases as it requires considerable human effort to label each sample correctly. In many cases like medical imaging, there is a small amount of labelled dataset along with large amount of unlabelled samples. In this research, we implement an Active learning algorithm which can help in increasing performance of deep learning models by using large amount of available unlabelled dataset. We propose a novel Active learning algorithm (Triplet AL) which uses a triplet network to select samples from unlabelled set for training classification model. Past active learning methods rely on classification model's final prediction scores as a measure of confidence for an unlabelled sample. We propose a more reliable confidence measure called Top-Two-Margin which is given by Triplet Network. We used STL-10 and CIFAR-10 dataset to test proposed algorithm. To test architectural independence of proposed algorithm, we tested proposed algorithm by using different model architectures for classification model. We compared results obtained using proposed method with past active learning methods. Proposed algorithm outperforms other active learning approaches we used to compare in our research.