The performances of different CNNs are assessed based on the classification accuracies that are observed on the validation set. The results are presented in Fig. 2. In this figure, the bars are ordered in the same way as columns in Table 3. For each considered CNN architecture, 3 CNN instances have been trained from scratch (each time an initialization of weights is random) and in Fig. 2, the bars represent the mean accuracies. Corresponding standard deviations are given by error segments. We have fixed a selection threshold accuracy of 97.5% on the validation set (shown by the dash-dotted horizontal line in Fig. 2), which corresponds to the accuracy of the Starting CNN on the validation set. This threshold is used to select CNN architectures in all 3 steps of the optimization procedure.