A New Multi-objective Optimization Model for Optimal Configuration of CNNs

In recent years, convolutional neural networks (CNNs) are compute-intensive learning models with growing applicability in a variety of real-world problems. However, determining an optimal CNN architecture is still unclear and needs mathematical development. Furthermore, most current CNN models are primarily crafted in a specific manner. This study introduces a novel multi-objective optimization model to find optimal hyperparameters related to CNN architecture. Our proposal simultaneously minimizes both cross-entropy loss and a related CNN complexity function. This model enables the fine control of each neuron’s contribution within the CNN’s fully connected layers through a set of associated decision variables. To solve the resulting model, we opt for the evolutionary multi-objective algorithm (NSGA-II). Extensive experiment on three benchmark datasets MNIST, SVHN, and NORB demonstrate the effectiveness of our proposed approach.