TSP: Task-Specific Pruning for Personalized Image Classification on Edge Devices

Abstract

Convolutional Neural Networks (CNNs) are extensively utilized for image classification. However, their high complexity hinders deployment on resource-constrained edge devices. Although existing pruning schemes could compress CNNs, they often rely on complete datasets and fail to account for the fact that practical edge tasks may not require handling all categories equally, leading to redundancies within pruned models. Moreover, training smaller task-specific models from scratch is both time-consuming and computationally expensive. To address these issues, we propose a Task-Specific Pruning (TSP) method that customizes pruning process to focus on categories pertinent to the specific task. Specifically, starting from a pre-pruned, full-scale model, TSP incorporates task-specific information into pruning process to identify and remove filters irrelevant to the targeted task, thereby achieving further compression. Experiments demonstrate that TSP could significantly reduce model size up to 8.7% on CIFAR-10 and 21.4% on CIFAR-100, while maintaining comparable accuracy to baseline methods.