Document Type
Honors Thesis
Abstract
MobileCLIP is a compact model that connects images and text, enabling it to perform tasks like image identification and question answering without needing to be retrained for each new task. Although it’s designed to be lightweight, it still runs slowly on regular computers without a powerful graphics card (GPU). This research focuses on making MobileCLIP run faster and smaller by using post-training quantization, which reduces the model’s precision after training without hurting performance. We combined several strategies: analyzing which parts of the model are more sensitive to changes, applying targeted adjustments to its structure, and running everything using CPU-only tools. Our tests showed we could shrink the entire model by 53% (from 397MB down to 186MB) and cut its overall processing time by 5%, while keeping its accuracy nearly the same. This work shows that advanced AI models can run on everyday computers, opening the door to more accessible, energy-efficient AI.
Disciplines
Computer and Systems Architecture
Publication Date
Fall 2025
Faculty Mentor of Honors Project
Marnim Galib
License
This work is licensed under a Creative Commons Attribution 4.0 International License.
Recommended Citation
Rahman, Atiqur, "Advancing MobileCLIP Through Hybrid Quantization: A CPU-Focused Approach" (2025). 2025 Fall Honors Capstones Projects. 26.
https://mavmatrix.uta.edu/honors_fall2025/26