DEVELOPMENT AND OPTIMIZATION OF ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING ALGORITHMS

Abstract

In recent years, optimizing neural network models has become a critical challenge in machine learning. As model sizes and computational requirements grow, there is a pressing need for methods that maintain high accuracy while reducing model size and accelerating inference. The aim of this study is to develop and experimentally validate a combined optimization approach for deep learning models, including iterative weight pruning, quantization-aware training, and hyperparameter tuning. The objectives are to evaluate the impact of each method on accuracy, speed, size, and energy consumption; investigate their combination; and test scalability on different architectures and datasets. Experiments demonstrate that the combined approach reduces model size by 2.6–3.8 times on average, speeds up inference by 2–2.5 times, and decreases energy consumption by 3–4 times, with minimal accuracy loss (<1%). Scalability was confirmed on ResNet18 and CIFAR-10. Final accuracy remains at 99–99.5% of the original, and error patterns are preserved. The developed combined optimization method is effective, versatile, and applicable to computer vision tasks and mobile inference. Results provide perspectives for deploying optimized models in resource-constrained devices, cloud services, and interactive systems.