A few recent contributions to sequential data acquisition
Building sample-efficient models is key when dealing with a limited budget, or when acquiring a new data point is expensive (money, time, computational cost...). In the context of sequential model building, the next data point should then be chosen very carefully. What's the next point to add to the training set ? What's the next experiment to run ? What question should be asked to an expert ? These questions are the focus of the interconnected fields of active learning, Bayesian experimental design, and Bayesian optimization, which all provide principled strategies for sequential data acquisition. In this talk, I'll detail a few recent contributions to these fields. In the first part, I'll introduce a new active learning strategy when the goal is not merely model accuracy, but is to take into account a down-the-line decision-making problem. In the second part, I'll discuss the problem of bringing human experts into Bayesian optimization, with the goal of integrating their knowledge to speed up the optimization.