Data-Aware Training Quality Monitoring and Certification for Deep Learning

Farhang Yeganegi, Arian Eamaz, Mojtaba Soltanalian

Keywords: Data-aware training quality monitoring, suboptimal training detection, YES bounds

Abstract: Deep learning models achieve remarkable representation power, yet their black-box nature raises concerns in high-stakes applications. While generalization analysis is well studied, less attention has been given to certifying the training process itself. We introduce the YES training quality bounds, a framework for real-time, data-aware certification and monitoring of neural network training. The bounds assess data utilization and optimization dynamics, revealing issues such as loss plateaus in suboptimal regions. Validated on synthetic and real data across classification and denoising tasks, YES bounds reliably certify training quality. Integrated with a color-coded cloud monitoring system, they provide a practical tool for real-time evaluation, setting a new standard for training quality assurance in deep learning.

Submission Number: 85