OLML：在线机器学习系统研究

OLML（Online Machine Learning）是一类新型的数据库管理系统，将机器学习模型训练任务看作是一种 “复杂数据分析”任务，支持用户的即时（ad hoc）式的在线模型训练和部署。OLML系统以机器学习所涉及的数据和模型为管理对象，通过“复用”等手段，降低模型训练和部署的成本、缩短模型训练和部署时间以满足在线模型训练的要求，从而实现“在线”学习的目的。.研究内容包括系统体系架构设计、机器学习描述语言MLDL的设计、训练数据之间以及训练数据与模型之间的语义关联、自动生成候选训练计划、自动选择最优训练计划、评测基准的构建、以及原型系统研制等。关注的核心科学问题是：1）如何实现“小时级”的模型学习与部署 2）如何有效地管理训练数据 3）什么是“好的”OLML系统. 通过研究形成完整的OLML的概念体系和原型系统，做出原创性成果。

OLML（Online Machine Learning）is a new category of database management systems. It treats model training tasks in machine learning as a kind of "complex data analysis" tasks, and supports ad hoc model training and deployment in one-hour level. For this purpose of online machine learning, OLML decouples machine learning process into a serious of steps and a set of data sets, from raw data set to final models. By reusing data sets and models among different training tasks and managing all objects in machine learning efficiently, we hope to upgrade the efficiency of OLML system significantly..The research topics include OLML system architecture, machine learning declarative language design, extraction of semantic relations among training data and between training data and model, automatical generation of training plans, optimization of training plans, benchmark evaluation, and prototype system development, etc..We focus on three technique challenges:1) How to implement an OLML system that can learn and deploy a model in one-hourlevel. 2) How to manage training data and models efficiently 3) What is the standard of a "good" OLML system. We hope to build a complete set of concepts for OLML and conduct leading research in this new field in the world.