Primal and dual first order methods for SVM: applications to driver monitoring

Machine learning tools are become recently very popular for solving real applications from many areas. Most of the learning problems are formulated as optimization problems with simple objective function but large number of constraints of order the number of training data. When considering the dual formulation, usually the objective function is difficult to minimize but the constraints are simple. One relevant application that fits into this pattern is the support vector machine (SVM). A popular approach for solving the primal SVM problem is based on first order methods due to their superior empirical performance. When considering the dual SVM formulation, which has simple constraints, coordinate descent schemes are typically the method of choice in practice due to their cheap iteration. In this paper we present a comparative study of several first order methods for solving primal or dual SVM problems. Numerical evidence on support vector machine classification for automatic detection of driver fatigue supports the effectiveness of such first order methods in real-world problems.