Enhancing CNN structure and learning through NSGA-II-based multi-objective optimization

In recent years, the advancement of convolutional neural networks (CNNs) has been driven by the pursuit of higher classification accuracy in image tasks. However, achieving optimal performance often requires extensive manual design, incorporating domain-specific knowledge and problem-understanding. This approach often results in highly complex network architectures, overlooking the potential drawbacks of such complexity. To this end, we propose MOGA-CNN, a Multi-Objective Genetic Algorithm for CNN structure that treats the CNN architecture design as a bi-objective optimization problem. MOGA-CNN aims to simultaneously optimize classification accuracy and minimize computational complexity, as measured by the number of learnable parameters. We employ the NSGA-II algorithm to effectively explore the trade-offs between these two conflicting objectives. The main contribution of this paper is the development of an encoding mechanism that captures the essential hyperparameters that influence CNN architecture, including the fully connected layer. To evaluate the effectiveness of our proposed algorithm, we conducted extensive experiments on four datasets, comparing its performance against other state-of-the-art methods. The results consistently demonstrate that our approach achieves satisfactory results when compared to these approaches.