There is no significant difference between the accuracies of the Starting CNN, the CNN A and the CNN B. All of them show the validation results which are very close to the threshold accuracy (with respect to standard deviations). The accuracy of the CNN C is significantly lower than accuracies of the first 3 networks: about 1.5% decrease of accuracy is observed on the validation set. Therefore, as the objective is to reduce the complexity of the Starting CNN while preserving its performance, the CNN B is selected after this first optimization step.