Learning Graph Representations for Intelligent Visual Computing

新兴的智能视觉计算领域寻求利用计算机视觉、几何处理和深度学习的最新进展来智能分析图像、视频和3D形状等视觉数据。深度神经网络在分析视觉数据方面的快速应用在很大程度上归功于负担得起的计算硬件、开源软件和大量训练数据的访问。在本课程中，我们将研究图形信号处理、光谱几何和深度学习如何相互作用来确定视觉数据识别系统的整体性能，以及如何最好地联合优化这些组件。为了给出三个相互交织的问题的答案，本课程将特别强调3D形状识别：如何学习稳健的形状感知表示；将学习转移到其他任务和图形结构数据的模式的最佳方法是什么；以及学习的表示的最佳用途是什么。这个项目的目标是在设计和开发用于学习深层图形表示的区分和生成模型方面研究有前途的想法，并在此方向上建立我们现有的一些工作。具体地说，我们的目标是：(I)通过引入来自图形信号处理和谱几何的更多合理的概念来形式化用于图形结构数据的深度神经网络的设计和分析；(Ii)研究所开发的算法的效率和可扩展性；(Iii)为图像和3D对象识别开发理论严谨和计算可行的方法；(Iv)探索所开发的方法在不同视觉数据领域的可转移性；以及(V)开发开发的解决方案的新的视觉计算应用。除了具有重要的理论意义外，拟议的研究计划的结果将是对更好地理解视觉数据的最先进技术的重要贡献，并将对包括神经成像和推荐系统在内的各种视觉计算应用产生影响。该项目不仅将通过促进加拿大本已蓬勃发展的人工智能研究和技术，为加拿大的社会和经济发展做出贡献，而且还将有助于培训一批可供加拿大工业界、学术界、政府机构和私人组织使用的技能人才。

The newly emerging field of intelligent visual computing seeks to leverage recent advances in computer vision, geometry processing and deep learning to intelligently analyze visual data such as images, videos and 3D shapes. The swift uptake of deep neural networks in analyzing visual data is largely attributed to a combination of affordable computing hardware, open source software, and access to large amounts of training data.In this program, we will study how graph signal processing, spectral geometry and deep learning interact to determine overall visual data recognition system performance, and how best to jointly optimize these components. A particular emphasis will be placed on 3D shape recognition in an effort to provide answers to three intertwined problems: how to learn robust shape-aware representations; what is the best way to transfer learning to other tasks and modalities for graph-structured data; and what are the best uses for the learned representations.The objective of this program is to research promising ideas in the design and development of discriminative and generative models for learning deep graph representations, and also to build on some of our existing work along this direction. In particular, we aim to: (i) Formalize both the design and analysis of deep neural networks for graph-structured data by introducing more sound concepts from graph signal processing and spectral geometry; (ii) Investigate the efficiency and scalability of the developed algorithms; (iii) Develop theoretically rigorous and computationally feasible methodologies for image and 3D object recognition; (iv) Explore the transferability of the developed approaches to different visual data domains; and (v) Exploit novel visual computing applications of the developed solutions.In addition to having significant theoretical implications, the outcome of the proposed research program will be important contributions to the state-of-the-art techniques geared toward a better understanding of visual data, and will also have an impact on a variety of visual computing applications, including neuroimaging and recommender systems. Not only would this program contribute to the social and economic development of Canada by boosting its already flourishing AI research and technology, but it would also contribute toward the training of a number of skilled personnel available to Canadian industry, academia, government agencies, and private organizations.