Interactive Learning: Combining Machine Learning Strategies with Humans in the Loop
Postdoctoral Fellow, Machine Learning Department, Carnegie Mellon University
Thursday, November 3, 2011 3:00pm
Newell Simon Hall 3305
People learn by interacting with their teachers. Why not machines? This talk presents recent machine learning approaches to natural language systems, which interact with human annotators in an attempt to learn more quickly and economically. These systems combine multiple learning strategies, for example: incorporating domain knowledge (taking advice in the form of human-provided rules), active learning (asking “questions” of human annotators in the form of unlabeled data instances or rules), and semi-supervised learning (attempting to “teach itself” by extrapolating what has been learned onto abundant, unlabeled examples). Empirical results from user experiments show that these systems are superior to their state-of-the-art “passive” learning counterparts. Interestingly, these experiments also provide some initial insights into human annotator behavior as well, suggesting ways in which human factors can and should be taken into account in interactive learning. I will point the way forward by outlining several relevant open questions for research in machine learning (both theory and practice) as well as in human-computer interaction.