Quantization Alters AI Reasoning Despite Preserving Accuracy.

Post-training quantization (PTQ) is a common technique to deploy deep learning models on resource-constrained edge devices by compressing high-precision weights into lower-precision representations. While past research has shown that PTQ largely preserves classification accuracy, its impact on a model's internal reasoning and interpretability has remained an open question. This study systematically evaluated how static PTQ affects the explainability of five widely used Convolutional Neural Network (CNN) architectures (VGG19, ResNet18, EfficientNet-B0, DenseNet161, MobileNetV2) at INT8 and INT4 precision. Using a dual interpretability framework combining Grad-CAM for spatial attention and LIME for input-level feature attribution, researchers compared full-precision and quantized models on binary classification datasets. The findings indicate that classification accuracy is not a reliable proxy for interpretability stability under reduced precision. For instance, DenseNet161 maintained strong feature consistency, while EfficientNet-B0, despite competitive accuracy, showed substantial degradation in input-level feature attribution. This research underscores that architecture selection is as critical as the quantization strategy when deploying models with high interpretability requirements, as the underlying "reasons" for predictions can change even if the predictions themselves remain accurate.