Export Reviews, Discussions, Author Feedback and Meta-Reviews

The paper describes a family of multi-class SVMs parameterised by a single parameter p. The family contains the Crammer & Singer SVM as a special case. Corresponding generalization bounds are derived, which for certain parameter choices have a logarithmic dependency on the number of classes. The latter has - to my knowledge - not been shown for any multi-class classifier. The corresponding learning machine is derived including its Fenchel dual problem.