Genetic Algorithm-Based Pruning for Efficient DenseNet Architectures

CNNs have shown remarkable performance on a variety of computer vision problems. However, CNN-based models require a lot of computational resources, which have limitations of resource-constrained environments. To address this problem, various lightweight techniques have been developed, such as pruning of network structures. This paper employed a genetic algorithm (GA) to implement pruning with various pruning rates, aiming for the efficient DenseNet. We optimized the dense connectivity pattern of DenseNet-BC ($k=12$) using a GA-based pruning method with multi-dimensional encoding scheme. We demonstrate that the proposed method can perform similarly with fewer parameters than the baseline model.