Using more than 4 particles is not so easy Training time increases Numerical fails are probable Classification is generally worse So combine parts intelligently (#(1/loss)**3#)