Multi-prototype Support Vector Machines

We extend multiclass SVM to multiple prototypes per class. This allows to combine several simple models (e.g. linear models) in a principled way to get much more expressive decision functions. For this framework, we give a compact constrained quadratic formulation and we suggest an ef£cient algorithm for its optimization that guarantees a local minimum of the objective function. An annealed process is also proposed that helps to escape from local minima. Finally, we report experiments where the performance obtained is almost comparable to that obtained by state-of-art kernel based methods but with a signi£cant reduction (of one or two orders) in response time.