Abstract: In any type of decision making, the ability to predict the outcomes of different options will determine the quality of the decision made. However, traditional predictive models trained with supervised learning capture the biases of the underlying training data and are generally not robust to spurious correlations. By instead learning the underlying causal relationships, causal models are able to bypass these limitations and therefore lend themselves well to decision optimization. In this talk we will look at the general setting in which causal ML is required. We will then look closer at what our team at Microsoft Research focuses on, decision making, and how we are approaching real problems in the area. Finally, we will look at our recent work in the area, deep learning based end-to-end methods, while looking towards what’s ahead.

Bio: Agrin works as a senior research engineer at Microsoft Research in fellow RPL alumni Cheng Zhang’s team that pursues research in the direction of deep causal learning.

Organizer: Marcus Klasson & Ruibo Tu