Frugal AI: Sustainable Deep Learning Compression via Pruning and Mixed-Precision QAT

Abstract

Deep learning models frequently require substantial computational resources, leading to high operational carbon emissions. This paper applies the principles of sobriété numérique (digital sobriety) to optimize a Convolutional Neural Network (CNN) trained on the CIFAR-10 dataset. By integrating iterative magnitude pruning (70% sparsity) with mixed-precision Quantization-Aware Training (QAT), the final model achieves a 3.73x reduction in file size (from 2085.93 KB to 559.53 KB) with a 0.67% drop in test accuracy (76.15% to 75.48%). The end-to-end optimization pipeline emitted 12.006 g CO2e, demonstrating that severe model compression is viable under strict environmental constraints.