Toward building a cost-effective and robust machine learning system

Artificial intelligence and machine learning (ML) are changing the world. They are more and more a part of our lives, empowering smartphones, online shopping, business analytics, etc. The technology behind the success of modern ML, deep learning, although having excellent predictive power, often requires very high cost to develop, from collecting data to training and evaluating models. In this talk, we will discuss several aspects of building a cost-effective ML system, from using active learning to reduce data collection cost to applying continual and transfer learning to reduce model training cost. We will also focus on the problem of transferability estimation and model selection that can be used for efficiently selecting the best source model to transfer to a target task, thus enabling the development of an ML model for the target task using only a small amount of data and computation time.