3. T. Darrell, A. Quattoni, M. Collins: Multi-task learning of Predictive Structures for Visual Category Learning
The transfer of knowledge from previously learned classes
to novel
tasks is a core challenge in large scale visual category learning. By
exploiting predictive structures found in previous tasks, which
presumably have relatively large amounts of training data, performance
can be improved on novel problems with sparsely labeled data. We
explore the learning of predictive structures based both on experience
from the previous categories, and on so-called auxiliary problems
constructed to provide additional tasks [Ando and Zhang 2005]. We show that
such techniques can improve the performance of an image
category classifier trained with just a labeled few examples.