Compressing Convolutional Neural Networks for Offline Educational Deployment An Empirical Study of Pruning and INT8 Quantization

Access to artificial intelligence tools in rural and low-resource educational settings is fundamentally constrained by hardware: most schools have, at best, low-end smartphones or shared desktop computers, and rarely have internet connectivity reliable enough for cloud-based inference. This paper presents a fully reproducible empirical study of two standard model-compression techniques—magnitude-based weight pruning and post-training 8-bit (INT8) quantization—applied to a compact convolutional neural network (CNN) trained for handwritten digit recognition on the MNIST dataset. Using a controlled experimental design, we measure the accuracy, model size, and compression ratio of four deployment configurations. These results provide a transparent, low-cost, and entirely reproducible reference point for deploying handwriting-recognition tools on inexpensive edge hardware without requiring GPUs or cloud services. Methodology This paper presents a fully reproducible empirical study of two standard model-compression techniques: Magnitude-based weight pruning (targeting 70% structural sparsity) Post-training 8-bit (INT8) quantization These techniques were applied to a compact convolutional neural network (CNN) trained for handwritten digit recognition on the MNIST dataset. Using a strictly controlled experimental design, an independently cloned copy of a trained baseline model was pruned and fine-tuned to isolate the exact effects of compression without shared-state artifacts. Key Findings The empirical results isolate the distinct architectural and physical impacts of both techniques under standard deployment tooling (TensorFlow Lite): The Baseline: The FP32 CNN achieved 98.78% test accuracy with a 224,892-byte physical footprint. The Quantization Impact: INT8 quantization reduced the footprint to 61,800 bytes (a 3.64x reduction) with a negligible 0.09 percentage-point shift in accuracy. The Pruning Impact: Iterative pruning to 70% structural sparsity matched the baseline almost perfectly at 98.77%, but left the physical flatbuffer file size unchanged due to standard dense storage formatting. The Combined Approach: Combining both techniques perfectly restored accuracy to 98.78% while maintaining the 3.64x physical compression. Conclusion These results offer practical, evidence-based guidance for institutions in low-resource settings considering deploying offline AI tools on inexpensive edge hardware, demonstrating that INT8 quantization acts as a near-cost-free first step for deployment.