Data enhancement is to enhance the generalization ability of the model, the nature of it and some other methods such as dropout, what is the difference between weight decay?

CodePudding user response:

(1) the weight decay, dropout, stochastic methods such as the depth, is specifically designed to limit the effective capacity of the model, is used to reduce the fitting, this kind is explicitly regularization method, studies have shown that this kind of method can improve the generalization ability, but it is not necessary, and ability is limited, and the factors such as parameters are highly dependent on the network structure,
(2) data increase would not lower the network capacity, also do not increase the computational complexity and quantities, is implicitly structured method, it makes more sense in practical application, so we often say that the data first,