CRITERIA FOR THE EFFICIENCY AND QUALITY OF NEURAL NETWORKS

Abstract

The article examines the criteria for evaluating the efficiency and quality of artificial neural networks (ANNs), emphasizing the importance of balancing predictive performance with computational feasibility. In the era of rapidly growing data volumes and increasing model complexity, the need for systematic approaches to assessing both quality and efficiency becomes critical. The study provides a comprehensive classification of metrics into three major groups: quality metrics, efficiency metrics, and generalization properties. Quality criteria such as accuracy, precision, recall, F1-score, AUC-ROC, and regression-based measures (MSE, MAE, RMSE, R², MAPE) are analyzed as primary indicators of predictive reliability. At the same time, efficiency is measured through computational costs (training time, inference time, FLOPs, memory footprint, energy consumption), structural parameters (number of layers and parameters, compression potential), and practical adaptability. Generalization ability is addressed through overfitting and underfitting analysis, validation and test errors, cross-validation, and the bias–variance trade-off. Furthermore, the paper highlights the importance of robustness and reliability criteria, including sensitivity to noise, adversarial resistance, reproducibility, and stability across datasets. Integral evaluation approaches, such as weighted score, quality-to-cost ratio, and sustainability indices, are proposed as tools for holistic assessment of ANN performance in real-world environments. The practical significance of the study lies in enabling informed decision-making in architecture design and hyperparameter optimization, supporting efficient deployment of neural networks in domains ranging from real-time embedded systems and IoT devices to large-scale industrial and medical applications. The findings emphasize that a balanced multi-criteria framework not only ensures predictive accuracy but also promotes resource efficiency, scalability, and long-term reliability of AI solutions, thereby contributing to sustainable and context-aware artificial intelligence development.