Shiwali Mohan

Speaker

Dr. Shiwali Mohan is a senior member of research staff at Xerox PARC. She studies the design and analysis of collaborative human-AI systems. Her research brings together methods from artificial intelligence (AI) and machine learning (ML) with insights from human-centered sciences to design systems that can collaborate with humans effectively. Her research has had an inter-disciplinary impact and has been published at venues for research on AI, human cognition, cognitive systems, human-computer interaction (HCI), medical informatics, AI for social good, and robotics.

Dr. Mohan has served as a principal investigator for grants from DARPA and AFOSR. In addition, she has been a key personnel on grants from ARPA-E, DOE, and other government agencies. She has also led Xerox research on interactive intelligent systems. Her recent work on intelligent coaching systems for health behavior change is one of the first demonstrations of long-term, interactive, intelligent behavior that was evaluated with human trainees in ecological settings. During her doctoral research, she led the development of Rosie - a complex agent with end-to-end behavior including natural language processing, flexible dialog, computer vision, action control, and interactive learning. This research paved the way for a new challenge problem for integrated AI systems research - Interactive Task Learning. With John Laird, she won the Blue Sky Award at AAAI 2018 for proposing a novel framework that integrates statistical learning with cognitive reasoning for complex intelligent behavior. She is an emerging leader of the AI community. She served as the chair for AAAI Doctoral Consortium at AAAI 2020 & 2021 and co-chaired the annual meeting of the Advances in Cognitive Systems (ACS) community in 2020.

Dr. Mohan received her B.E. in Instrumentation and Control Engineering from Netaji Subhas Institute of Technology, Delhi University. She received her M.S. and Ph.D. degrees in computer science from the University of Michigan, Ann Arbor with a focus on artificial intelligence.

Speaker Links: Website - Google Scholar - Linkedin

Abstract

It is expected that general-purpose, autonomous machines will become pervasive in domestic, public, and industrial spaces within the next decade. They will assist humans in a variety of activities including doing household tasks and collaborating on the assembly line. Personal robots, along with other intelligent agents such as smart homes and cars, will add tremendously to the quality of human life. They will offer persons with impairments more independence, help older adults with their daily chores, transport people and goods, and perform search and rescue in environments that are too dangerous for humans. Several challenges have to be addressed to make progress towards this vision. Each home, office, or assembly line is organized differently. Users will want the agents to perform a variety of tasks and will have different preferences. Customizing every agent for its deployment environment and user preferences is resource-intensive and costly. One approach to this challenge is designing a generally intelligent robot that can adapt to the user requirement on its own instead of relying on dedicated programming. This approach requires that the agent be an efficient online learner. It needs to learn object recognition, semantic organization and categorization, spatial relations, and tasks from experiences in its environment and exploit this knowledge immediately for performance. Learning through self-directed exploration alone can be challenging and slow, requiring numerous interactions with the environment. With my colleagues, I study how human feedback, guidance, and structure can be used to reduce the complexity of robot learning. Our goal is to develop agents whose behavior can be easily extended by human users using natural interactions. This research agenda is a fascinating and gratifying mix of insights from various scientific disciplines including cognitive science, human-computer interaction, machine learning, reasoning & inference, and robotics. In this talk, I will introduce the research agenda - Interactive Task Learning - and present the latest advances.

Papers covered during the talk

• Mohan, S., Klenk, M., Shreve, M., Evans, K., Ang, A., Maxwell, J. (2020). Characterizing an Analogical Concept Memory for Architectures Implementing the Common Model of Cognition pdf
• Ramaraj, P., Ortiz Jr, C. L., Klenk, M., Mohan, S.(2021). Unpacking Human Teachers’ Intentions For Natural Interactive Task Learning pdf
• Interactive Task Learning link
• The Soar Cognitive Architecture link