Novel Post-Training Structure-Agnostic Weight Pruning Technique for Deep Neural Networks

Deep neural networks (DNNs) have shown ex-ceptional performance in various domains, leading to their widespread adoption. However, the necessity to deploy DNNs on resource-constrained devices calls for improved model efficiency. Accordingly, DNN pruning has emerged as a critical technique in the field of machine learning, offering significant improve-ments in computational efficiency and model simplicity. This paper introduces an innovative post-training pruning approach for DNNs without any retraining, that utilizes multi-objective optimization to achieve substantial sparsity rates while preserving significant accuracy levels. The proposed method transforms the post-training weight pruning challenge into a two-variable, bi-objective optimization problem. The optimizer finds the optimal minimum and maximum threshold values through optimization, effectively converting the real-valued weights between these two thresholds to zero. The task and model-independency of the proposed framework make it applicable across various models, tasks, and datasets without constraints on the number of weights. The approach provides a decision-maker, where users can select the best strategy within resource constraints to achieve their desired accuracy. In order to assess our method, we evaluated the pruning of RESNET50 model on CIFAR10 and CIFAR100 benchmark datasets. In CIFAR10, by reducing 70% of the weights within the optimal threshold values, the network's accuracy only decreases by 0.1. Similarly, in CIFAR100, an appropriate weight range was selected, resulting in a 65 % reduction in weights while maintaining a negligible 0.1 decrease in network accuracy. This demonstrates the effectiveness of the optimization in achieving significant model size reduction without compromising performance on large DNNs.