Human-In-The-Loop Machine Learning for Health

Abstract

Many health sciences related tasks involve limited amounts of data : pharmacogenetics, drug design, clinical trial design, etc. These are typically not well-suited for modern AI algorithms. To circumvent this issue, human feedback can be leveraged, provided it is accounted for in the statistical models, and that efficient strategies facilitate its collection. The field of research spanned by these topics is called Human-In-The-Loop Machine Learning (HITL-ML). The aim of this presentation is to illustrate several contributions in this area. As a gentle introduction, I will begin with a study from colleagues in my team, incorporating expert feedback to improve genomic-based predictions. Next, I will introduce an algorithm dealing with unreliable information sources in the framework of Bayesian Optimization, thus allowing to leverage human experts of unknown relevance. I will conclude with some work in progress, regarding the application of HITL-ML tools in the context of de novo drug design.