Model optimization by destructive methods: A case study using structured pruning in image classification models
Year: 2024
Type: conference-paper
Abstract: The Deep Neural Network architectures’ advances offer innovativeand effective solutions to various complex challenges. In order toimprove the models’ effectiveness based on these architectures’ requirements,the model hyperparameter optimization is an essentialproject stage. Despite this, the existing optimization techniquesdemand a high computational cost when directly applied to thestate-of-the-art most complex architectures. In this sense, this studyproposes a method for hyperparameter optimization leveraginglow computational cost for Artificial Intelligence models throughstructured pruning techniques. For the purpose of verifying anddemonstrating the method, three main hypotheses are investigatedthroughout the implementation of a case study, which consists ofthe models’ fine-tuning and optimization for three-dimensionalgeometric shapes classification on real-world object images. Theprocess includes hyperparameter optimization for pruned models,considering the posterior retraining, evaluation, analysis and comparisonbetween the performance and efficiency of the originalmodels. Finally, the results were promising, indicating an improvementthat reaches up to 10.58% Precision by just focusing on themodels’ learning rates optimization.
Source: Anais do Computer on the Beach
Institution: Universidade Federal do Pará
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4399324376