A scalable channel pruning framework with computation-adaptive artificial bee colony

Channel pruning, known for its hardware-friendly characteristics of structural pruning, has been widely adopted to slim deep neural networks. However, the increasingly complex coupled structures in neural networks pose new challenges to channel pruning methods. First, most existing channel pruning approaches exhibit customized characteristics, making them difficult to adapt to the multiple new neural networks derived from coupled structures embedded in neural networks. Second, the intricate inter-layer dependencies within coupled structures are crucial for enhancing model performance, requiring careful pruning to prevent severe performance degradation in the pruned model. Meanwhile, continuously evolving coupled structures further exacerbate these challenges. To achieve high-ratio pruning of deep neural networks while preserving model performance and ensuring the scalability of the pruning method, this paper proposes a computation-adaptive artificial bee colony approach for scalable channel pruning. First, this paper proposes a structure mapping strategy to construct a scalable search space, which decouples complex networks by establishing different mapping relationships between network architectures and search structures. Then, this paper designs a computation-adaptive artificial bee colony algorithm that reduces fitness estimation bias by utilizing local evaluation consistency among sampling networks. Finally, a cell-level search space construction method is introduced to exponentially reduce the search space. Experimental results demonstrate the superiority of the proposed method on computer vision tasks. Additionally, ablation studies and visualizations verify the effectiveness of the proposed approach.