SiamALNet: A Semi-supervised Siamese Neural Network with Active Learning Approach for Auto-Labeling

Deep learning algorithms have been successfully used for classification, and it provides state-of-the-art accuracy when trained on a large amount of labeled datasets. Acquiring the labels consumes huge costs and time. So, there is always a need for auto-labeling of the unlabeled data, which can reduce the labeling cost to a marginal amount. This study’s proposed model, ‘Semi-supervised Siamese Neural Network with Active Learning (SiamALNet),’ combines semi-supervised learning (SSL) with consistency-based training, pseudo-labeling approach, and active learning. Most confident and uncertain data points of the unlabeled pool are annotated using a Siamese neural network with consistency-based training and active learning, respectively. The confident labels and manually annotated labels are further used with labeled data to retrain the model. This process continues until all the unlabeled data are annotated. The auto-labeling performance of the SiamALNet is tested on different datasets, namely ZF internal dataset, Human Activity Recognition (HAR), MNIST, and CIFAR-10. The proposed model achieves more than 99% auto-labeling accuracy for most datasets. In this work, the importance of consistency-based training and active learning is also studied by comparing the accuracy of the SSL model with the proposed SiamALNet model. The output of SiamALNet is embeddings, making it ideal for similarity search in unlabeled data.