Learning structure from document bases; the vast majority of available electronic documents reside in unstructured document bases. Hypertext remains woefully inadequate: it is labor-intensive, static, one-directional, and limited to the perspective of the hypertext author. And when we move beyond the traditional text document to audio, video, and other forms of recorded information, even the rudimentary benefits of hypertext are missing.
Can we automate the extraction of relationships between documents to generate a dynamic structure that lets users navigate and manipulate unstructured document bases at the conceptual level?
Alternative models of robotics (a.k.a "moving bits, not bolts"); traditional robots require the ability to sense, compute, act on the world and, optionally, communicate. By letting communication substitute for some of these other abilities, can we build smaller, simpler, and yet more powerful robotic agents? What uses might we have for a "symbot" whose only ability to act is by communicating with its environment? What about a thin-client robot ("thin-bot"), which does no computation of its own, but acts as the eyes and hands of one or more remote computer?
Active learning; many learning problems in robotics, pattern recognition, and information retrieval afford chances for the learner to select or influence the training data it receives. How should it select training data to learn most quickly, for the least cost?