Analysis and Design of Algorithms for Interactive Machine Learning

The traditional role of the human operator in machine learning problems is that of a batch labeler, whose work is done before the learning even begins. However, there is an important class of problems in which the human is interacting directly with the learning algorithm as it learns. Canonical problem scenarios which fall into this space include active learning, interactive clustering, query by selection, learning to rank, and others. Such problems are characterized by three main factors:

1. the algorithm requires input from the human during training, in the form of labels, feedback, parameter guidance, etc.
2. the user cannot express an explicit loss function to optimize, either because it is impractical to label a large training set or because they can only express implicit preferences.
3. the stopping criterion is performance that's "good enough" in the eyes of the user.

The goal of this workshop is to focus on the machine learning techniques that apply to these problems, both in terms of surveying the major paradigms and sharing information about new work in this area. Through a combination of invited talks, discussions, and posters, we hope to gain a better understanding of the available algorithms and best practices for this space, as well as their inherent limitations.

For more information, see http://research.microsoft.com/~sumitb/adaiml09