The Impact of Model Capacity on Performance via Weight Pruning: A Study Using the Food-11 Dataset

Abstract

Deep learning models, particularly convolutional neural networks, have exhibited outstanding performance across a wide array of complex tasks and applications. Although CNNs have proven invaluable in numerous image-related tasks, they come with high computational demands and memory requirements, limiting their deployment in resource-constrained environments, such as mobile devices and embedded systems. Addressing this challenge, model size reduction, and weight pruning have emerged as essential strategies for optimizing model performance and efficiency. This study investigates the relevance of employing high-capacity models and the impact of weight pruning in the context of a specific dataset, the Food-11 dataset and the DenseNet121-based model. Weight pruning is the focal point of this study, with a specific focus on percentage-based pruning. Results show a trade-off between model accuracy and size as pruning percentages increase. While some accuracy loss is observed, the reduction in model size and evaluation time is substantial, making it a compelling strategy, especially in resourceconstrained scenarios. This study underscores the potential benefits of weight pruning as a means to enhance the efficiency and applicability of deep learning models. It highlights the importance of finding the right balance between model compression and performance, ultimately enabling the deployment of deep learning solutions in real-world applications with limited computational resources.