Learning to Reuse Visual Knowledge

The central question in my talk is how existing knowledge, in the form of available labeled datasets, can be (re-)used for solving a new (and possibly) unrelated image classification task. This brings together two of my recent research directions, which I'll discuss both. First, I'll present some recent works in zero-shot learning, where we use ImageNet objects and semantic embeddings for various classification tasks. Second, I'll present our work on active-learning. To re-use existing knowledge we propose to use zero-shot classifiers as prior information to guide the learning process by linking the new task to the existing labels. The work discussed in this talk has been published at ACM MM, CVPR, ECCV, and ICCV.