As a result, the proposed 3-step methodology has proved to be very efficient in minimizing the Starting CNN for the gender prediction from face images. However, the problem of choosing an optimal CNN architecture for a specific problem remains an open subject. In this paper, we do not pretend to answer it in a general case, as the main goal of our study is designing an optimized and efficient CNN model for gender recognition from face images. Nevertheless, the proposed empirical CNN optimization methodology can be easily adapted to any problem of interest. For that, we suggest to start from an established and well-known CNN architecture (as it is done with the Starting CNN in this work) and progressively minimize it using the proposed 3-step methodology. Moreover, our approach can be combined with other optimization strategies (e.g. [7,9]) to minimize running time and required memory of a CNN model.