Large Margin Learning with High Order Loss Functions