EAGER: Large-Scale Distributed Learning of Noisy Labels for Images and Video

该项目开发了从带有噪声标签的图像和视频中学习的算法。在线免费提供的大量图像和视频给机器学习和计算机视觉研究社区带来了前所未有的挑战。它们也为解决图像理解中的人机语义差距以及彻底改变我们索引、检索和与图像和视频交互的方式带来了巨大的机会和潜力。不准确的标签和错误的标签是图像和视频数据集的常见问题。有噪声的标签会给现有的学习算法带来问题。这个项目可以对其他大数据问题产生广泛的影响。该项目通过培训学生、确保代表性不足群体的广泛参与以及向公众推广，与教育相结合。本研究探索带有噪声标签的大规模图像和视频的分布式学习方法。PI研究了平滑和非平滑正则化项的损失函数的学习问题，并相应地开发了新的分布式学习算法，该算法能够利用大量太大而无法放入单个机器的图像。该研究在图像和视频分析以及计算机视觉应用方面具有巨大的潜力。具体而言，本研究强调算法和理论两个方面：(1)开发基于分布式学习的方法来优化和学习噪声标签；(2)研究保证收敛、收敛速度和可扩展性等问题。这项工作为许多领域的新发现提供了广泛适用于许多经济、医学和科学上重要的大规模数据集的新方法。

This project develops algorithms for learning from images and video with noisy labels. The overwhelming amounts of images and video freely available online present unprecedented challenges for machine learning and computer vision research communities. They also bring tremendous opportunities and great potentials for addressing human-machine semantic gaps in image understanding and for revolutionizing our ways to index, retrieve, and interact with images and video. Inaccurate labels and mislabeled data are common problems for image and video datasets. Noisy labels would cause problems with the existing learning algorithms. This project can have broad impacts on other big data problem. The project is integrated with education by training students, ensuring broad participation of underrepresented groups, and outreaching general public.This research exploring distributed learning methods for large-scale images and video with noisy labels. The PI investigates the learning problem of loss functions with both smooth and non-smooth regularization terms, and accordingly develops new distributed learning algorithms that are capable of leveraging the abundance of images that are too large to fit into a single machine. The research has an immense potential in image and video analysis, and computer vision applications. Specifically, this research emphasizes both algorithmic and theoretic aspects by (1) developing distributed learning based approaches for optimization and learning of noisy labels; and (2) investigating issues such as guaranteed convergence, convergence rate, and scalability. This work provides new methods that are widely applicable to many economically, medically and scientifically important large-scale datasets for novel discoveries across many domains.