Efficient Optimization of Multi-class Support Vector Machines with MSVMpack

In the field of machine learning, multi-class support vector machines (M-SVMs) are state-of-the-art classifiers with training algorithms that amount to convex quadratic programs. However, solving these quadratic programs in practice is a complex task that typically cannot be assigned to a general purpose solver. The paper describes the main features of an efficient solver for M-SVMs, as implemented in the MSVMpack software. The latest additions to this software are also highlighted and a few numerical experiments are presented to assess its efficiency.