As machine learning (ML) advances, it is likely to have a big impact on any domain with lots of data and easily measurable objectives. This encompasses a large swath of daily life, but it is currently unclear whether ML can help in matters that require substantial thought. For example, as individuals, can ML help us reflect on our personal goals and on how to achieve them? As a society, can it help us make progress on the problems that we collectively face, such as finding better ways to help those most in need?
I will consider this question in the context of building automated systems that use back-and-forth dialog to help their users think through tricky questions. I will review existing approaches to dialog automation and the challenges they face, and describe an alternative that is intended to be more suitable for conversations that require substantial thought. Instead of directly learning to predict the next sentence given the conversation so far, we learn "cognitive actions" on a workspace that is shared with human contributors. Humans help out where our algorithms aren't yet capable enough. To incentivize such help, we build a market around crowdsourced contributions. The result is a Dialog Market: a mechanism for creating high-quality conversations that resolve vague questions.
About the Speaker: Andreas Stuhlmüller received his B.Sc. in Cognitive Science from the University of Osnabrück in 2009 and his Ph.D. in Brain and Cognitiv Sciences from MIT in 2015. At MIT, he was part of Josh Tenenbaum's Computational Cognitive Science group, where he worked on the design and implementation of probabilistic programming languages and on their application to cognitive modeling. He did postdoctoral work at Stanford University and is currently a researcher in Noah Goodman's Computation & Cognition lab.
