Layer-Wise Filter Thresholding Based CNN Pruning for Efficient IoT Edge Implementations

This paper presents a novel approach for efficiently running convolutional neural networks (CNNs) on Internet of Things (IoT) edge devices. The proposed method utilizes threshold-based pruning to optimize pre-trained CNN models, enabling inference on resource-constrained IoT and edge devices. The pruning thresholds for each layer are iteratively adjusted using a range-based threshold pruning technique. The pre-trained network evaluates the accuracy of the pruned model and dynamically adjusts the pruning thresholds to maximize accuracy. The effectiveness of the proposed approach is validated on the widely-used LeNet benchmark network, with MNIST, Fashion-MNIST, and SVHN datasets. Our experimental results show that for the MNIST dataset, we can prune 62–64% of weights for an accuracy loss of 1–4%. Similarly, for Fashion-MNIST, we can prune around 64% for an accuracy loss of around 2.92%, and for the SVHN dataset, we can prune around 55% of weights for an accuracy loss of 1.7% on average, saving resources.