Adaptive Lifelong Learning and Open Vocabulary Object Retrieval
What's the difference between coffee mugs displayed on internet images and the one I have on my office desk? When can we finally teach robots to bring us our coffee mug in the morning? What have coffee mugs to do with visual bird recognition?
The talk will not be about coffee mugs the whole time but I will give an overview of my research in the area of lifelong visual learning. In particular, I try to cover the aspects of domain adaptation, object retrieval, and part detector discovery. For domain adaptation, I will focus on an approach that allows for joint learning of category-invariant transformations and classifier parameters. In contrast, open vocabulary object retrieval deals with finding the right object for a given user query ("Robot, please give me the teapot.") and I show how simple natural language processing can be combined with state-of-the-art vision techniques to solve this task. In the last minutes of the presentation, I will talk about how implicit object part detectors can be discovered in convolutional networks trained on ImageNet.
Ort: Seminarraum des Institut für Neuro- und Bioinformatik
