Department of Computer Engineering

Training a deep learning model with supervised learning methods requires large amounts of labeled data, which is difficult to obtain for many fields. We propose a data programming approach that enables automatic labeling by defining programmable functions. We apply this weakly supervised learning-based approach to visual relationship classification task, which aims to identify relationships between objects in images. Our results demonstrate that the ground truth labels and the performance of fully supervised learning can be approximated by creating only five functions involving different weak sources.