2017年12月20日下午15点，应广州大学经济与统计学院和岭南统计科学研究中心的邀请，普渡大学统计系Chuanhai Liu教授在行政东前座412会议室作了题为“Automated Building of Deep-Learning Models: an Expand-and-Reduce Method”的讲座——暨“羊城讲坛”第二十一讲，旨在进一步提高年轻学者及研究生对研究的理解。此次讲座由系主任张兴发主持，相关专业的师生参加了此次讲座。本报告讲述数据分析需要计算集型的数据驱动方法，因为它比以往更需要探索和建模复杂的数据结构。由于其能够灵活地表达复杂的本地和全球数据结构，“多层网络”或“深度学习”模型实际上可能非常有用。然而，这样的模型的架构配置和参数优化在统计和计算上都是非常具有挑战性的。为了克服这个困难，我们提出了一种Expand-and-Reduce方法来自动构建深度学习模型。

 

 

摘要：Data analysis demands computer-intensive data-driven methods more than ever for exploring and modeling complex data structures. Because of its capability and flexibility to represent complex local and global data structures, ``multi-level nets'' or Deep Learning models can be practically very useful. However, architecture configuration and parameter optimization of such models are extremely challenging, both statistically and computationally. To overcome the difficulty, we propose an Expand-and-Reduce method for automated building of Deep-Learning models. The purpose of the method is three-fold: 1) it generates models for Artificial Intelligence-type of applications, 2) it can be used to do confirmatory-type analysis for investigating prior knowledge-based construction of network structures, and 3) it provides as a tool for in-depth investigation and understanding of data from scientific inference perspective. With simple examples, we show that our proposed method is promising to serve its purpose.