Reinforced SVM method and memorization mechanisms

The paper is devoted to two problems: (1) reinforcement of SVM algorithms, and (2) justification memorization mechanisms for generalization. Current algorithm was designed the case when risk set nonnegative slack variables defined by l1 norm. In this paper, along with that classical norm, we consider risks l2 norm l∞ Using these norms, formulate several modifications existing show resulting modified algorithms can improve (sometimes significantly) classification performance. Generalization ability learning usually explained arguments involving uniform convergence empirical losses corresponding expected over a given functions. However, bounds losses, VC theory also provides relative convergence. These lead more accurate estimate loss. Advanced methods estimating error have leverage bounds, which support training data memorization, which, as demonstrates,