Towards Efficient Deep Neural Networks via Post-Training Quantization

This thesis advances post-training quantization (PTQ) for efficient deployment of state-of-the-art deep learning models while maintaining their performance. PTQ offers significant advantages, including data privacy preservation, minimal computational overhead, and rapid deployment. However, it faces fundamental challenges that become increasingly severe at lower bit-widths, particularly overfitting to small calibration datasets and model-specific complexities arising from different network architectures. This research addresses these challenges across Convolutional Neural Networks, Diffusion Models, and Large Language Models. Our progression from convolutional neural networks (CNNs), diffusion models (DMs), and large language models (LLMs) demonstrates how dedicated solutions address architecture-specific challenges, with each contribution adapting to the specific requirements of different model families.