Human-in-the-Loop AI for Autonomous Problem Solving (or why humans ruin the best things)

Abstract

With the coming of the new AI revolution, the expectations on autonomous systems and their decision-making algorithms have increased manifold. Traditional (symbolic) AI techniques – which are crippled by a dependence on complete and correct models of the world, yet can offer much more transparency and customizability than the black-box neural techniques that are de rigueur today – must be dragged kicking and screaming into this brave new world. In this talk, I offer the example of autonomous technical support bots – a problem domain that is as relevant today in 2017 as it was in 1992 when the pioneering work on Unix Softbots began – as an example research testbed. I will show how attempting to solve this problem in today’s data-rich yet model-poor world brings together various AI techniques such as reinforcement learning, automated planning, information extraction, and natural language processing. By relying on signals of causality, correlation, and preference that are hidden in human-generated data, our techniques show a path towards bridging the gap between brittle models that only work in limited cases, and the reality of today’s unstructured world.

Bio

Kartik Talamadupula is a Research Staff Member at IBM Research AI, in the T.J. Watson Research Center in New York. He graduated with a Ph.D. in Computer Science from Arizona State University in 2014, advised by Prof. Subbarao Kambhampati. His thesis work focused on understanding, analyzing, and extending the role that automated planners can play in integrated AI and Robotics systems that interact directly and cooperatively with humans. His work has been published at AAAI, ICAPS, ICLR, HRI, IROS, EMNLP, IAAI, ACS, ICWSM, CIKM, WebScience, and HCOMP; and presented at various international workshops and systems demonstrations. He was part of the team that won the Best Demo award at AAAI 2018 and the Best Demo Runner-Up award at ICAPS 2018; and the Best Paper award at the MRQA Workshop at ACL 2018. He has been a recipient of the Science Foundation Arizona Fellowship; ASU's University Graduate Fellowship; and received a letter of commendation from the University President's Office for contribution to ASU's research and academic enterprise. He graduated summa cum laude with a Bachelors degree in Computer Science from Arizona State University, and was recognized with the Distinguished Undergraduate Senior award, and a CRA Outstanding Undergraduate Researcher Award Honorable Mention.