Building a stable and performant Distributed System requires careful planning, design, and implementation. The real challenge with distributed system, however, starts after system is built and rolled out. End users and system administrators spend a lot of time, far more than development time, using and operating the system. Thus providing ease-of-use and easy-to-operate for both users and system administrators is a critical must-have feature for any distributed system.
In this talk, I will present Nuage, LinkedIn’s private cloud management portal, built to bring ease-of-use and operability to LinkedIn’s Distributed Data systems. We will go over why LinkedIn made investment in Cloud Management, list of features and benefits Nuage brings to application developers and platform administrators, and the future investments we are making.
In his talk, Xiaobing will cover below topics:
Machine learning, especially deep learning, has gained more and more attention over the past years. However, resources wise it is more efficient for some companies to leverage the existing models and APIs built by big IT companies which have large datasets and huge machine power. In 2017 and forward, we expect to see a high rise of usage for Cloud-based AI.
Change Luo will cover several new developments related to Cloud-based AI at Google: vision, speech, translation as well as Q&A. He will also demo a state of the art example leveraging all those techniques together and build your intelligent conversational app (chatbot) with little effort and prior knowledge. It uses computer vision to detect the landmark inside a picture, a natural language understanding engine that can understand what users ask and provide the most relevant information. Then it leverages the language API to generate speech that almost indifferentiable from normal person. Finally it translates with Cloud Translation API.
In this talk I will present gradient boosted decision tables (BDTs). I will present novel algorithms to fit decision tables and discuss why decision tables are better weak learner in the gradient boosting framework. In addition, I will talk about efficient data structures to represent decision tables and a novel fast algorithm to improve the scoring efficiency for boosted ensemble of decision tables. In the end, I will also discuss our successful deployment boosted decision tables to LinkedIn news feed system that achieved significant lift on key metrics. This work has been published in KDD'17.