Deadline-Driven CNN Model Based on Residual Quantization
Ali Haider Alvi, Yoichi Tomioka, Yuichi Okuyama, Jungpil Shin · 2025
In hard real-time systems with strict timing constraints, completing inference within the given deadline is crucial. Model compression methods, such as quantization, have been proposed to reduce inference time. However, these methods do not guarantee that inference will meet the deadline and may degrade accuracy due to aggressive compression. This paper presents a novel deadline-driven model that combines residual quantization with dynamic skipping of residual components to meet hard deadlines while maintaining high accuracy. When tested on the CIFAR-10 dataset using the ResNet-20 architecture, the model achieves a 0% violation rate for timing constraints and delivers accuracy comparable to models that miss deadlines. Compared to non-deadline-driven models, it provides a flexible solution for real-time deployment in hardware-constrained environments, ensuring reliable performance under varying timing demands and resource availability.