Efficient Machine Learning

Deep learning has significantly advanced computer vision but remains limited by heavy computation, slow inference, and privacy concerns. This thesis introduces resource-efficient solutions: (1) three knowledge-distillation methods—ℓ₁-regularized teacher training, teacher fine-tuning, and language-guided re-ranking—to close the student–teacher performance gap; (2) a parallelized diffusion sampling algorithm that eliminates sequential dependencies, accelerating high-quality image synthesis; and (3) a training-free machine unlearning mechanism that removes targeted concepts by excising corresponding subspaces in the weight space, rendering models blind to specified content. Collectively, these techniques reduce model size, speed up inference, and enforce data privacy while maintaining state-of-the-art accuracy.