Ian Fasel
Assistant Research Professor
Department of Computer Science
University of Arizona
Gould-Simpson Building
1040 E. 4th Street
Tucson, AZ 85721
ianfasel at cs dot arizona dot edu
Biography
I attended the University of Texas at Austin for a B.S. in electrical engineering and a B.A. in Plan II Honors, then went to UCSD for a Ph.D. in Cognitive Science in Javier Movellan’s Machine Perception Lab, and then returned to UT Austin for a postdoc with Peter Stone in the AI Lab in the Department of Computer Sciences. Throughout 2007 and 2008 I spent several months at the University of Osaka working with Professors Hiroshi Ishiguro and Minoru Asada in the Asada Synergistic Intelligence project. In 2009 I joined the University of Arizona as an assistant research professor.
My research is in cognitive science, machine perception, artificial intelligence, and what one might call “developmental robotics”. This means developing real physical agents that have to solve some of the same problems that the developing human brain has to solve. Usually this involves learning sensory skills (such as real-time detection of objects, faces, and facial expressions, real-time classification of non-verbal speech, or classification of touch in robot skin) from very little supervision, or from unsupervised, active exploration. In one recent project we learned real-time object detectors for many object categories from only tens of unsegmented images. In another project, we built a robot baby that used contingencies between it’s vocalizations and simple audio “loudness” events to drive learning of a visual object detector. Amazingly, although this robot had never been told anything specifically about faces, it learned a fast face detector from only a few minutes worth of interaction with human caregivers.
A current focus is expanding on the idea of information maximization (InfoMax) as an intrinsic utility in POMDPs for coordinating teams of robots, and allowing human teammates to easily guide them by giving them “informational” goals and feedback. In another project, we are developing agents that can learn from non-expert human teachers through natural instruction. A key ingredient is to discover structure in the world and across lessons in an unsupervised manner so that the agent can learn not only what the teacher was explicitly trying to tell it, but also everything else in the world that makes the spare, often very ambiguous human instruction still enough to learn complex concepts and tasks.