At the 3rd AutoML Challenge, GrandMasters, a team of AI engineers from Inspur Group, Beijing University of Posts and Telecommunications (BUPT), and Central South University (CSU), ranked 3rd among more than three hundred participating teams. Audodidact.ai, a team from a startup company, won first place in the competition, and Meta_Learners, a team led by Professor Zhu Wenwu from the Department of Computer Science and Technology, Tsinghua University, placed second. The challenge, with the theme “Lifelong ML with Concept Drift,” came ahead of NeurIPS 2018 (i.e. Conference on Neural Information Processing Systems), held in Montreal, Canada from December 2nd to 7th. Nearly 10,000 participants were expected to attend the conference, making it the world’s largest AI academic event. The conference will feature an awards ceremony for the winners of the AutoML Challenge, with the top three teams delivering presentations.
The AutoML Challenge is the world’s top event in automated machine learning, launched at NeurIPS this year for the first time. Numerous high-caliber talents from leading universities and companies participated, including Tsinghua University, Peking University, Massachusetts Institute of Technology, Texas A&M University, Microsoft, Tencent, Alibaba, Autodidact.ai, etc.
As one of the hottest fields of artificial intelligence research, Automated Machine Learning (AutoML) helps automate the workflow of machine learning, reducing the dependence on algorithm developers and shortening the time for model development. However in many real-world applications, AutoML can face numerous challenges. Batches of data may be arriving daily, weekly, monthly, or yearly, for instance, and the data distributions are changing relatively slowly over time. This presents a continuous learning or Lifelong Machine Learning challenge for an AutoML system. Typical learning problems of this kind include online advertising, recommendation, sentiment analysis, fraud detection, spam filtering, traffic monitoring, econometrics, patient monitoring, and climate monitoring. This year’s challenge focuses on Lifelong machine learning. Participants are invited to design computer programs capable of automated lifelong machine learning with requirements of algorithm versatility, adaptability, robustness and running time.
By creating new data features, GrandMasters increased accuracy by 27% based on the original model. By fine-tuning the lightGBM model, the team further enhanced the model performance by 10%. Inspur’s AI engineers independently designed a feature selection algorithm, shortening the model training time by 30% through reducing the dimensions of data features. The engineers also proposed a sliding-window approach to data training to strike a balance between the training time and the calculation accuracy. The team improved the performance and accuracy of the model as well as reducing the training time, making their program a leading contender in the competition.
“AutoML lowers the threshold of machine learning and helps promote machine learning applications and industrial transformation,” said Liu Jun, General Manager of AI & HPC, Inspur. “As a professional AI computing power provider, Inspur is committed to the development and application of AI technologies. This award shows that Inspur’s AI team ranks among the top in AutoML algorithms. Inspur will be inspired to further develop AI technologies, drive the application and transformation of AI computing with our partners from home and abroad.”