Overparametrized deep network predict well despite the lack of an explicit complexity control during training such as an explicit regularization term. For exponential-type loss functions, we solve this puzzle by showing an effective regularization effect of gradient descent in terms of the normalized weights that are relevant for classification.