Large-scale learning has evolved significantly in the last decade. While significant progress has been made by focusing on how to build flexible and powerful input representations, i.e. image features, we are interested in the complementary direction of learning a structure in the output space. Most of the prior approaches build upon a large-margin binary classification framework and assume that the output space consists of an arbitrary finite set of labels or class identifiers. We consider the case where elements of the label set are structured objects which have their own identifiers. Intuitively, rather than learning if an image contains a zebra, we learn that the object contained in an image is a black and white, has stripes and looks like a horse. This way of embedding labels in a vector space is an effective tool to model latent relationships between classes. In addition to providing an alternative for large-scale image classification, these embedding methods have shown to be effective when labeled training data is scarce.
In this tutorial, we first provide essential theoretical foundations of the structural embedding framework, followed by discussing practical applications of such embeddings in computer vision and machine learning literature. Our tutorial on embeddings will also contain a practical session where the participants will have the opportunity to experiment with embeddings applied to zero-shot image classification problem. Finally, we provide a bridging course on the connections between embeddings and distance metric learning where the goal is to replace a standard distance function, such as the Euclidean distance, with a learned distance function
References:
-
Large Margin Methods for Structured and Interdependent Output Variables, Tsochantaridis etal, JMLR 2005
-
Label Embedding Trees for Large Multi-Class Tasks, Bengio etal, NIPS 2010
-
Label-Embedding for Image Classification, Akata etal, TPAMI 2015
-
Distance metric learning, with application to clustering with side-information, Xing etal, NIPS 2002
-
Metric Learning by Collapsing Classes, Globerson & Roweiss, NIPS 2006
-
Distance Metric Learning for Large Margin Classification, Weinberger & Saul, JMLR 2009
-
TagProp: Discriminative Metric Learning in Nearest Neighbor Models for Image Auto-Annotation, Guillaumin et al, ICCV 2009
-
Learning to Rank with (a Lot of) Word Features, Bai et al., JMLR 2010
-
Distance-Based Image Classification: Generalizing to new classes at near-zero cost, Mensink et al., TPAMI 2013
Tentative Program:
09:00-09:45 Embeddings
Talk by Zeynep Akata
09:45-10:00 Break
10:00-11:00 Applications of Embeddings
Talk by Zeynep Akata (Max Planck Institute for Informatics, Saarbrucken)
Practical Session by Yongqin Xian (Max Planck Institute for Informatics, Saarbrucken)
11:00-11:15 Break
11:15-12:00 Metric Learning
Talk by Thomas Mensink (University of Amsterdam, Netherlands)
