Level Supervision

Level supervision in machine learning focuses on training models with varying degrees of labeled data, ranging from full annotations to only image- or video-level labels. Current research emphasizes developing techniques to effectively leverage these weaker supervision signals, often employing pseudo-labeling, contrastive learning, and multi-task learning within architectures like neural radiance fields, transformers, and convolutional neural networks. This area is significant because it allows for training powerful models with less expensive and more readily available data, impacting diverse applications such as video analysis, medical image segmentation, and 3D reconstruction.