Compression of Deep-Learning Models Through Global Weight Pruning Using Alternating Direction Method of Multipliers

Abstract

Abstract Deep learning has shown excellent performance in numerous machine-learning tasks, but one practical obstacle in deep learning is that the amount of computation and required memory is huge. Model compression, especially in deep learning, is very useful because it saves memory and reduces storage size while maintaining model performance. Model compression in a layered network structure aims to reduce the number of edges by pruning weights that are deemed unnecessary during the calculation. However, existing weight pruning methods perform a layer-by-layer reduction, which requires a predefined removal-ratio constraint for each layer. Layer-by-layer removal ratios must be structurally specified depending on the task, causing a sharp increase in the training time due to a large number of tuning parameters. Thus, such a layer-by-layer strategy is hardly feasible for deep layered models. Our proposed method aims to perform weight pruning in a deep layered network, while producing similar performance, by setting a global removal ratio for the entire model without prior knowledge of the structural characteristics. Our experiments with the proposed method show reliable and high-quality performance, obviating layer-by-layer removal ratios. Furthermore, experiments with increasing layers yield a pattern in the pruned weights that could provide an insight into the layers’ structural importance. The experiment with the LeNet-5 model using MNIST data results in a higher compression ratio of 98.8% for the proposed method, outperforming existing pruning algorithms. In the Resnet-56 experiment, the performance change according to removal ratios of 10–90% is investigated, and a higher removal ratio is achieved compared to other tested models. We also demonstrate the effectiveness of the proposed method with YOLOv4, a real-life object-detection model requiring substantial computation.

Keywords

• Network compression
• Weight pruning
• Non-convex optimization
• Parallel computing