1: \begin{abstract}

2: We present a dual subspace ascent algorithm for support vector machine

3: training that respects a budget constraint limiting the number of

4: support vectors. Budget methods are effective for reducing the training

5: time of kernel SVM while retaining high accuracy. To date, budget training

6: is available only for primal (SGD-based) solvers. Dual subspace ascent

7: methods like sequential minimal optimization are attractive for their

8: good adaptation to the problem structure, their fast convergence rate,

9: and their practical speed. By incorporating a budget constraint into a

10: dual algorithm, our method enjoys the best of both worlds. We demonstrate

11: considerable speed-ups over primal budget training methods.

12: \end{abstract}

13: