Computational Relevance of Model Pruning and Quantization for Low-Powered AI
Ashmit Mandal, Nafisa Hasan, Arumai Jain, Jnyana Ranjan Mohanty, Vandana Sharma · 2025
This study investigates how to make machine learning models more efficient for low-power devices by simplifying their structure and lowering size. Six models were evaluated: feed-forward neural networks (FNN), convolutional neural networks (CNN, notably VGG), decision trees, random forests, support vector machines (SVM), and auto encoders. Each was evaluated in its original form, after pruning and quantization, with a focus on model size, accuracy, and training time (as a measure of energy use). The results indicated that pruning for parameter like model size is greatly minimized with Decision Trees reducing by 95%, while it is observed quantization increases efficiency even more. In case of parameter, accuracy, it declined by about 7%. Results show that VGG model retain more accuracy than others after quantization. Pruning also increased training time, particularly for VGG and SVM models. This research thus provides insights into the trade-offs between model complexity, accuracy, and efficiency, guiding the selection for suitable models in resource-bounded environments.