Quantizing Deep Learning Vision Models - A Systematic Approach

As deep learning models continue to grow in complexity and size, their deployment on resource-constrained environments such as embedded systems and edge devices becomes increasingly challenging. Quantization has emerged as a key technique to reduce model size and improve inference efficiency by reducing the numerical precision of weights and activations. However, quantizing models without compromising accuracy remains non-trivial due to issues such as outlier sensitivity, activation range variability, and layer-specific behaviors. This paper presents a systematic and statistically grounded methodology for post-training quantization of deep learning vision models. Using AlexNet as a case study, the proposed workflow involves analyzing hyperparameter distributions, applying non-parametric statistical tests to guide quantization granularity, and visually inspecting parameter behavior across layers. The approach helps identify optimal clipping strategies and quantization configurations tailored to model architecture, enabling effective deployment with minimal performance degradation.