Lower the barrier of machine learning : meta learning for transfer learning and autoML

Abstract

Recent years, machine learning becomes the major methodology to develop artificial intelligence applications, due to the trend of using big data in machine learning. However, traditional machine learning may have three barriers: lack of data, poor feature quality, and less data scientists. In this thesis, we focus on how to lower the barrier of machine learning. We propose to use meta learning methodology to solve these problems. Specifically, meta learning can be applied to solve the transfer learning and AutoML problems. Transfer learning can be used to weaken the impact of small data and poor feature problems, and AutoML can be used to solve the problem that there are not enough data scientists and then we may use normal engineers to build up AI systems. As the result, the three main barriers have been lowered correspondingly. We designed several new algorithms to solve the data, feature and model tuning problems, and showed advantages on many empirical studies. As meta learning may rely on auxiliary data from other sources, we found that it may lead to privacy problem. To solve this problem and make meta learning better applied, we design a new privacy-preserving learning algorithm. In this algorithm, we show how to learn from data without accessing any privacy information.

Date of Award	2020
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology