Empowering Deep Learning with Graphs
Author | : Jiaxuan You (Machine learning researcher) |
Publisher | : |
Total Pages | : |
Release | : 2021 |
ISBN-10 | : OCLC:1268333482 |
ISBN-13 | : |
Rating | : 4/5 ( Downloads) |
Download or read book Empowering Deep Learning with Graphs written by Jiaxuan You (Machine learning researcher) and published by . This book was released on 2021 with total page pages. Available in PDF, EPUB and Kindle. Book excerpt: Deep learning has reshaped the research and applications in artificial intelligence. Modern deep learning models are primarily designed for regular-structured data, such as sequences and images. These models are built for tasks that take these regular-structured data as the input (e.g., classification, regression), as the output (e.g., generation), or as the structural prior (e.g., neural architecture design). However, not all forms of data are regular-structured. One notable example is graph-structured data, a general and powerful data structure that represents entities and their relationships in a succinct form. While graph-structured data is ubiquitous throughout the natural and social sciences, its discrete and non-i.i.d. nature brings unique challenges to modern deep learning models. In this thesis, we aim to empower deep learning with graph-structured data, by facilitating deep learning models to take graphs as the input, the output, and the prior. My research in these three directions has opened new frontiers for deep learning research: (1) Learning from graphs with deep learning. We develop expressive and effective deep learning methods that can take graphs as the input, which promotes the learning and understanding of graphs. (2) Generation of graphs with deep learning. We formulate the generation process of graphs using deep learning models, which advances the discovery and design of graphs. (3) Graph as the prior for deep learning. We discover that graph structure can serve as a powerful prior for neural architectures and machine learning tasks, which opens a new direction for the design and understanding of deep learning. Finally, we discuss the wide applications of the above-mentioned techniques, including recommender systems, drug discovery, neural architecture design, and missing data imputation.