2008 ICT Workshop on Intelligent Lifelong Learning Companions

In his novel The Diamond Age [2], Neil Stephenson tells the story of a young girl named Nell is given a copy of an interactive book called Young Lady's Illustrated Primer. The Primer stays with her, teaches her how to become a master engineer and how to gain confidence, as well as survival and leadership skills.  What would it take to build something as elaborate as the Primer? What advances have been made and need to be made to make something like this a reality? Clearly, vast amounts of knowledge and impressive reasoning capabilities would be needed. It would need to understand human learning, elicit individual needs and desires, monitor learning and growth, learn individual traits, provide emotional support, maintain personal relationships, be perceived as indispensable, and much much more.

The Primer would require some level of understanding of how humans acquire and learn new knowledge. Research in the learning sciences has revealed much about these cognitive processes and how to promote effective learning. Also, more and more emphasis is being placed on informal and self-directed learning, rendering many older notions of learning (e.g., "blank slate" instructionism) obsolete. Obviously, the Primer would also require pedagogical knowledge about how to explain and teach. Advances in intelligent tutoring systems have produced a body of AI techniques that feed into this need. Algorithms exist for assessment of learner activities, monitoring and tracking of learning, provision of feedback, individualization of learning experiences, and more. Further, work on embodied conversational agents and rapport has shown that virtual humans are able to form meaningful relationships with real humans, when they behave appropriately. The vision of a lifelong learning companion - one that stays with a learner over an extended period, monitors growth, gives suggestions, and motivates the learner, all with an understanding of the bigger picture - is arguably within our grasp.

How broad and deep of a problem is this for AI? Inspired in part by Stephenson's book, Marvin Minsky and his colleagues highlight a similar vision as an ideal application of commonsense reasoning research [1]. They, and the participants in the St. Thomas symposium, propose to develop a "personalized teaching machine that would adapt itself to someone's specific circumstances, difficulties, and needs." The system would need to "understand human goals", "understand human reasoning", and "converse in natural language." These clearly coincide with many ongoing AI research efforts, but also suggest a new way to frame our understanding of how these efforts may ultimately merge in important ways.

Questions and Goals

Few would argue that the problem of building an intelligent lifelong learning companion is a highly interdisciplinary one. A primary goal of this workshop is to bring together researchers from both the psychology and learning side of the problem as well as technologists who focus on interactive learning environments and AI. Presentations will address fundamental questions regarding learning and memory over extended periods of time. Also, we hope to address questions about technological advances that have been developed and need to be developed in order to build lifelong learning companions.  Many more questions can be asked about this problem:

• What are the implications of research into long-term human memory? How "sturdy" are memories and learned skills?  How easily do we forget? What are the ingredients for improving retention and reducing memory decay?
• How do learners behave with respect to broader learning and achievement goals? What do counselors, advisors, and mentors do to motivate and support learners at this higher level?
• What advances in AI have been made that contribute directly to building a lifelong learning companion? What are the most crucial advances in artificial intelligence and the learning sciences that need to be made? 
• How is it that embodied agents are able to form meaningful bonds with human users? What are their properties and what will be required to make these connections more permanent and long-lasting?
• Are current learner and user modeling techniques suitable and scalable enough to handle the demands of lifetime modeling of learning and growth?  What sophistication in reasoning is missing in current approaches that can make extended modeling possible? How do human mentors track growth?
• What are the implications for intelligent tutoring systems that seek to guide learning beyond a single learning episode?