Dr. Thomas Dietterich
Oregon State University

Abstract

Over the past 25 years, machine learning research has made huge progress on the problem of supervised learning. This talk will argue that now is the time to consider two new directions.

The first direction, which is already being pursued by many groups, is Structural Supervised Learning in which the input instances are no longer independent but instead are related by some kind of sequential, spatial, or graphical structure. A variety of methods are being developed, including hidden Markov support vector machines, conditional random fields, and sliding window techniques.

The other new direction is Deployable Learning Systems. Today's learning systems are primarily operated offline by machine learning experts. They provide an excellent way of constructing certain kinds of AI systems (e.g., speech recognizers, handwriting recognizers, data mining systems, etc.). But it is rare to see learning systems that can be deployed in real applications in which learning takes place on-line and without expert intervention. Deployed learning systems must deal with such problems as changes in the number, quality, and semantics of input features, changes in the output classes, and changes in the underlying probability distribution of instances. There are also difficult software engineering issues that must be addressed in order to make learning systems maintainable after they are deployed.