 |
Adaptive Interfaces for Machine Learning |
|
 CSE Home |
AI Home |
About CSE |
Search |
Contact Info |
|
 |
Adaptive Interfaces for Machine LearningOverviewMachine learning is widely used in areas such as wrapper induction, credit approval, image analysis, data mining, and as the driving force of adaptation in intelligent user interfaces. A learning module can endow a system with the ability to extract valuable information from substantial amounts of data or to adapt to new circumstances. However, many traditional applications of machine learning involve their users --- the domain experts who provide and label training examples --- in only the most primitive ways. The systems operate on batches of prelabelled examples and neither provide nor request feedback during the learning process. Indeed, the only ``communication'' between the user and the learning system besides the provision of this batch of examples is to test the system on yet another batch of examples.Yet, experience and research shows that great benefits can be realized by machine learning applications which support a variety of interactions with the user. Work in active learning has shown that just allowing the system to pick which example to label next can greatly reduce the number of examples the user need consider. Our previous work on the SMARTedit text-editing system led us to the same conclusion: machine learning tasks can benefit from careful management of interactions with the user. The result is a system that requires fewer examples to learn a concept, acquires domain knowledge more effectively, and is more appropriate for users unfamiliar with machine learning. Our current work involves studying how to implement and manage novel interactions in intelligent interfaces. Although the work began with the SMARTedit system, we are currently investigating other machine learning domains which could benefit from enhanced interfaces. DIAManDOur first adaptive interface for machine learning called DIAManD sits atop the SMARTedit text-editing system. The interface exposes a small set of different interactions to the user and provides recommendations on which interaction to activate. DIAManD automatically initiates the best interaction if its confidence in that interaction is high enough. Thus, the system becomes an active participant in the discourse with the user. The user can still exercise full control (overriding the system's choice), but many users in our preliminary testing are delighted to receive feedback from the system and loath to ignore its advice.Our general approach and the DIAManD system in particular are discussed in the SMARTedit Talks Back paper available in the publications list below. The following picture shows a brief glimpse of DIAManD in action. Using a set of quantitative measures of the available interactions, DIAManD learns from the user's implicit and explicit preferences which interaction is the best to perform at a given point in the overall discourse. In this case, the user has already recorded a single iteration normally and recorded another one out of order (at SMARTedit/DIAManD's suggestion). SMARTedit then correctly completed the second example and now suggests allowing it to complete the third.
Publications
|
|||||||||||||||
|
Department of Computer Science & Engineering University of Washington Box 352350 Seattle, WA 98195-2350 (206) 543-1695 voice, (206) 543-2969 FAX [comments to wolf@cs.washington.edu] |
