Invited Tutorial: Human Assisted ML

Abstract

1 GOAL OF THE TUTORIAL In a several applications, societies often depend on the judgement of human experts to make consequential decisions, which have significant impacts [1, 3, 6, 7]. However, the timeliness and the quality of these decisions are often compromised by shortage of experts. For example, in any hospital with limited resources, patients may have to wait for a long time. On the other hand, simple mistakes made by common works in a factory can lead to a heavy loss. To overcome this challenge, there is a recent line of works which describe how machine and humans can work together to achieve superior performance to what they individually could have achieved. The goal of this tutorial is introduce these techniques as well as describing the consequence of the outcome of machine learning on society at large— the pitfalls and the possibly remedy to overcome them. More specifically, we aim to teach the possibility of designing human assisted machine learning models in two classical machine learning setups— supervised learning and sequential decision making, where the machine learning task is distributed between humans and machines in such a way, so that, together human and machine perform better than than what they would achieve on their own. Summarizing the goals of the tutorial evolve about the following problem: (1) Human assisted supervised learning, where given both human annotations and ground truth labels, we learn how to outsource samples to humans to achieve better performance. (2) Human assisted reinforcement learning, where, for example in autonomous driving, a machine wants to take quick control in an emergency situation, or vice-versa. For example, in collision avoidance systems in cars, the machine takes control during a potential threat, whether in other aeroplanes, humans takes control if there is a possibility of mid-air collisions.