Henry Lieberman and David Maulsby
If the software of the future is to provide intelligent assistance to users of computer interfaces, it will have to incorporate some learning capabilities. Software might remember user habits and preferences, adapt to the user's style of working, or be taught new capabilities by interaction with the user. Machine learning has been a problem long studied in artificial intelligence, but incorporating learning into an interactive interface agent brings up unique issues, such as visual representation, pointing input, and user feedback which are not usually considered in the context of traditional machine learning. The focus of this course will be on understanding what kind of machine learning capabilities will be useful in an interactive context, rather than the abstract characterization of learning algorithms. We will read and discuss current research papers, see and critique demonstrations of learning interfaces, and experiment with implementations of learning systems. Course will require student presentations of papers, one term paper exploring a theme of the course and/or reviewing readings, or a programming project.
Some of the questions we will explore include:
* What learning algorithms are best for software agents?
* What is the role of classical techniques such as explanation-based learning, statistical techniques, case-based reasoning?
* Is it possible for a learning system to be "too smart"?
* How does the learning system communicate what it has learned to the user?
* Is it better to learn from instruction or from examples, or both?
* Course introduction
* Examples and critiques of current learning systems
* Examples of PBD systems
* Intro to Machine Learning algorithms
* Instructibility in machine learning
* Interface issues in learning systems
* Integration of learning and interaction
* Future directions
* Projects and presentations