This is a project to provide improvements which may enable the emerging technology of sensor-driven models of human situation to move from laboratory demonstrations to practical widespread use in human-computer interaction systems. Interfaces as we see them today are largely static ? they act in the same way regardless of what human situations they are used in. In the past world of information workers in a fixed office setting, this was acceptable. However, as technological advances allow inexpensive computing devices to move into all the diverse settings of everyday life, this will no longer be appropriate. Our devices should adjust to the human context of each situation in a way that maximizes their ease of use and effectiveness in serving human needs across those varying situations. However, currently most interactive systems have no information about the human situations they are operating in, or the activities of the users they serve. This represents a significant barrier to creating interfaces which are more appropriate to the wider world they are now being placed in.
To overcome this basic obstacle, an emerging body of work has begun to develop techniques for modeling human situations and activities. Much of this work employs sensor-driven statistical models created with machine learning techniques. These models provide useful estimates of activities and situations. However, there are a number of serious practical barriers to widespread use of these techniques. This grant will support an aggressive body of work aimed at overcoming the most important of these barriers. One of the most important of them is the expense, difficulty, and disruptiveness to end-users of collecting sufficient training data to make these systems work. To address this problem, the project includes development and study of a collection of innovative techniques tuned to reducing the human costs associated with collecting the required training examples.
In addition, new techniques for internet-scale collection of training data will be developed. These techniques will enable a shift from the current practice of collecting a large amount of data from a few people to collection of a small amount of data from many people. Complementing this approach, new hierarchical modeling techniques will be developed which should allow a smooth and rapid transition from an initial generic model reflecting average behavior of many people, to models which are finely tuned to the particulars of an individual with minimal disruption for that end-user. Finally, the work to be funded here will explore the effectiveness of a variety of emerging advances in machine learning technology which have yet to be widely applied to human-computer interaction applications.
The impact of this research will go significantly beyond the specific issues it attacks because it is enabling in nature ? seeking to open up the possibility of what should be complete classes of new interface technology. In addition, this work is particularly important for some applications supporting special needs populations, and the project includes activities to enhance education.
