1: \begin{abstract}

2: 

3: 

4:     We present a novel column generation based boosting method for

5:     multi-class classification.

6:     Our multi-class boosting is formulated in a single optimization

7:     problem as in \cite{Shen2011Direct,cvpr2012Paisitkriangkrai}.

8:     Different from most existing multi-class boosting methods, which

9:     use the same set of weak learners for all the classes,

10:     we train class specified weak learners (i.e.,

11:     each class has a different set of weak learners).

12:     We show that using separate weak learner sets for each class leads

13:     to fast convergence,

14:     without introducing additional computational  overhead in the

15:     training  procedure. To further make the training more efficient

16:     and scalable,  we also

17:     propose a fast coordinate descent method for

18:     solving the optimization problem at each boosting iteration.

19:     The proposed coordinate descent method is conceptually simple

20:     and easy to implement in that it is a closed-form solution for

21:     each coordinate update.

22:     Experimental results on a variety of  datasets show that,

23:     compared to a range of existing multi-class boosting methods, the proposed

24:     method has much faster convergence rate and  better generalization

25:     performance in most cases. We also empirically show that the

26:     proposed fast coordinate descent algorithm needs less

27:     training time than the MultiBoost algorithm in Shen and Hao

28:     \cite{Shen2011Direct}.

29: 

30: 

31: 

32: 

33: \end{abstract}

34: