Etudiant : EL IDRISSI LAOUKILI MOHAMMED

Filière : Master Machine Learning Avancé et Intelligence Multimédia (MLAIM)

Encadrant : Pr. SABRI ABDELOUAHED

Annèe : 2026

Résumé : This master's thesis addresses the challenge of reducing the computational and memory requirements of deep learning models used in computer vision, motivated by the difficulty of deploying large, high-performing architectures on resource-constrained devices. The primary goal of this project is to identify the best possible trade-off between model complexity and practical deployment constraints — that is, to reduce model size, improve or preserve predictive performance, and enhance inference speed, all while respecting the limitations imposed by the target hardware. To this end, the project systematically studies and compares the main families of model optimization and compression techniques — namely pruning, quantization, knowledge distillation, low-rank factorization, weight sharing and tying, efficient architecture design (including neural architecture search and depthwise-separable convolutions), and adaptive computation through early-exit networks — in order to understand their impact on both predictive performance and computational efficiency. A controlled experimental framework was established using two representative vision architectures, a modified ResNet-50 convolutional network and a Vision Transformer, trained primarily on the CIFAR-10 dataset, with an additional evaluation on the HAM10000 medical imaging dataset to assess how the observed behaviors generalize to a different, imbalanced data domain. For each optimization family, several representative methods were implemented and evaluated under identical training protocols, datasets, and hardware conditions, and assessed using a consistent set of performance metrics (accuracy, precision, recall, F1-score) alongside efficiency metrics (number of parameters, model size, FLOPs, and inference latency and throughput on both CPU and GPU). Beyond quantitative measurement, the project analyzes how each compression technique affects the internal structure and behavior of the models, the trade-offs between compression level and accuracy degradation, and the conditions under which each approach is applicable, ultimately providing a comprehensive experimental foundation for balancing model size, accuracy, and speed under real-world hardware constraints in efficient deep learning.