Optimizing Neural Network Efficiency with Hybrid Magnitude-Based and Node Pruning for Energy-efficient Computing in IoT

Abstract

The Deep Neural Networks (DNN) are computationally intensive in terms of processing, energy and memory which becomes a bottleneck to run these models on edge devices. This research study provides a technique for pruning the neural networks to enhance the performance of deep learning models in IoT devices. The proposed method combines magnitude-based pruning, which merges insignificant weights based on their magnitude, with node pruning, which eliminates insignificant nodes based on their contribution to the network. The hybrid pruning technique is designed to be energy-efficient, reducing the computational overhead of deep learning models while maintaining their accuracy. The experimental results demonstrate that the proposed method can achieve significant reductions in model size and energy consumption with minimal loss in accuracy. The technique has the potential to enable the deployment of deep learning models on resource constrained IoT devices.