Metaheuristics for pruning convolutional neural networks: A comparative study

Due to their learning and adaptation capabilities, convolutional neural networks (CNNs) have demonstrated potential for machine learning and artificial intelligence applications. However, computationally complex CNN models hinder deployment on resource-constrained devices. To mitigate this, researchers have developed algorithms to reduce the computational requirements of such models. Network pruning, a model compression technique significantly reducing computing costs with negligible accuracy loss, has recently gained prominence. However, manual intervention is often required, posing challenges for users lacking domain knowledge. Automatic pruning using metaheuristic algorithms, a promising approach, has gained recognition in recent years but must address time-consuming model evaluations and searching large solution spaces. Various researchers have developed effective metaheuristic-based pruning approaches to address these limitations. However, a comprehensive comparative study on the effectiveness of different metaheuristics for network pruning is lacking in the literature. Moreover, the impact of objective function formulations and encoding schemes on the performance of metaheuristic-based pruning methods still needs to be explored. To bridge these gaps, this paper presents an extensive experimental study investigating the effect of various metaheuristics on network pruning, with a particular emphasis on objective formulations and encoding schemes. The experimental results provide a detailed analysis of the impact of metaheuristics, objective function formulations, and encoding schemes on network pruning.