Quantization-Induced Error Propagation: Activation Analysis Within the PEP-AI Framework for Sustainable High-Efficiency and Explainable AI

Abstract

Post-training quantization to INT8 enables energyefficient deployment of deep neural networks on edge devices, yet standard metrics such as mean Average Precision (mAP) fail to capture internal instabilities. We propose a Precise, Explainable, and Provable (PEP-AI) framework for validating quantized convolutional neural networks beyond output-level evaluation. We analyze intermediate activations across Faster R-CNN, YOLO, EfficientNet, and DenseNet using ONNX-based feature extraction. Our results show that despite stable detection performance, internal feature maps exhibit median relative errors of 4-6% in deeper layers, with cumulative deviations reaching up to 20%. We identify an attenuation effect where Max-Pooling and Batch Normalization suppress quantization noise, particularly in object regions, where errors correlate with detection failures. In addition, smaller activation deviations $(\sim 6{\%})$ may lead to missed detections in safety-critical scenarios such as autonomous driving. At the same time, INT8 quantization reduces model size by $\mathbf{6 7 {\%}}$ and inference time by $\mathbf{5 0. 2 {\%}}$, significantly improving energy efficiency and sustainability. We introduce a visual-statistical framework based on symmetric logarithmic and hexbin analysis to reveal hidden error distributions. Our findings highlight that feature-level validation is useful for energy-efficient AI deployment.