BIGDATA: IA: Distributed Semi-Supervised Training of Deep Models and Its Applications in Video Understanding

该项目研究了深度神经网络模型的半监督训练，采用分布式方式使用大规模的标记和未标记数据。深度神经网络最近在人工智能和相关科学领域得到了广泛部署，这在很大程度上要归功于标签良好的大数据集和提高的计算能力。然而，通常更大的未标记数据本身就被流行的深层模型的有监督训练排除在外。对许多新近的和新兴的数据集的未标记部分进行建模确实是极具挑战性的，这些数据集通常是非结构化的并且分布在网络的不同节点(例如，由摄像机网络捕获的视频)。该项目旨在探索如何有效地利用未标记和分布的数据来补充已标记数据的区分线索，以共同学习准确和健壮的深层模型。该研究将机器学习、计算机视觉和并行计算无缝结合，为研究生和本科生培养独特的跨学科研究和教育方案。尽管半监督学习和深度学习取得了进展，但这两者的融合大多是在单机环境下进行的小规模研究。然而，许多新的数据集很容易增长，超出了一台机器的计算甚至存储能力。因此，在并行计算平台上研究深度模型的半监督学习成为迫切需要。为了更好地应对这种情况，该项目开发了改进的网络架构以促进并行培训，并开发了自适应地在同步和异步模式之间切换以实现最佳效率的培训程序。其主要思想是将参数分布结合到神经网络中，并使用协变量匹配来协调不同机器上的网络行为。研究人员还探索了一种新的应用--视频序列的极端尺度时空动作标注，以对该项目中的算法和框架进行基准测试。

This project investigates semi-supervised training of deep neural network models using large-scale labeled and unlabeled data in a distributed fashion. Deep neural networks have recently been widely deployed in artificial intelligence and related scientific fields, largely attributing to well-labeled big datasets and improved computing capabilities. However, the unlabeled data, which is often bigger, is inherently ruled out by the prevailing supervised training of the deep models. It is indeed highly challenging to model the unlabeled parts of many recent and emerging datasets, which are often unstructured and distributed over different nodes of a network (e.g., the videos captured by a camera network). This project aims to explore how to effectively use the unlabeled and distributed data to complement the discriminative cues of the labeled data, to jointly learn accurate and robust deep models. The research seamlessly unifies machine learning, computer vision, and parallel computing, and fosters unique interdisciplinary research and education programs for the graduate and undergraduate students.Despite the progress on semi-supervised learning and deep learning, the confluence of these two is mostly studied on a small scale in single-machine environment. However, many new datasets easily grow beyond the computation or even storage capacity of a single machine. Hence, it becomes a pressing need to investigate the semi-supervised learning of deep models on parallel computing platforms. To better account for this scenario, this project develops improved network architectures to facilitate the parallel training, and the training procedure developed adaptively switches between synchronized and asynchronized modes for optimal efficiency. The main idea is to incorporate a parametric distribution to the neural network and use covariate matching to coordinate the network behaviors across different machines. The researchers also explore a novel application, extreme-scale spatial-temporal action annotation of video sequences, to benchmark the algorithms and frameworks in this project.