Optimization of Neural Networks for Edge AI Applications

Abstract

The rapid growth of Internet of Things (IoT) devices and real-time intelligent systems has increased the demand for efficient Artificial Intelligence (AI) models that can operate under constrained computational, memory, and energy resources. Edge AI applications require optimized neural networks that balance predictive performance with low latency and reduced hardware complexity. This study focuses on the optimization of neural networks for deployment in edge computing environments through model compression and architectural efficiency enhancement techniques. The methodology includes the application of pruning strategies to eliminate redundant network connections, quantization to reduce numerical precision of weights and activations, and knowledge distillation to transfer learning from large teacher models to lightweight student networks. Additionally, lightweight architectures such as MobileNet and depthwise separable convolution-based models are evaluated to improve computational efficiency. Experimental validation is performed on benchmark datasets under simulated edge device constraints to assess performance metrics including accuracy, inference time, memory usage, and energy consumption. The results demonstrate that optimized neural networks achieve substantial reductions in model size and computational cost while maintaining competitive accuracy levels compared to full-scale deep learning models. Quantization and pruning techniques significantly reduce inference latency, making real-time processing feasible on edge devices. Knowledge distillation further enhances performance stability in compact models by preserving essential feature representations. The study concludes that neural network optimization techniques play a crucial role in enabling scalable and efficient Edge AI systems.