Scholarlink

<p>Deploying trained Artificial Intelligence (AI) models on resource-constrained devices is challenging due to limitations in memory and computational power. To overcome these constraints while maintaining model accuracy, Pruned AI (PAI) models are essential. In response to these challenges, researchers have developed several compression models for Convolutional Neural Networks (CNNs) In this study, a firefly algorithm-based approach was used to minimize the CNN model’s size by choosing a portion of neural network filters and nodes in the hidden layers, keeping the model’s original accuracy for the task of image classification. The ResNet50, VGG16, VGG19 and SqueezeNet architectures were used for model reduction, demonstrating significant gains in reduced storage space without compromising accuracy. The study demonstrates that this approach can achieve an additional 25% reduction in storage space even for SqueezeNet, a model that is already compressed. Moreover, the model’s accuracy improved by 0.23% on the MNIST dataset after pruning with SqueezNet and ResNet50. Similarly, the accuracy also improved for the other pre-trained models like VGG19 1.7% on the MNIST dataset, 7.78% for the CIFAR 10 dataset, and 3.09% for the CIFAR 100 dataset. Finally, an explainable AI model such as Grad-CAM was implemented on uncompressed and compressed models to verify the results and achieve higher confidence.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FCDC: Firefly optimization algorithm based compression of deep CNN model with explainability

  • 期刊论文

  • 发表日期：

    2025年10月17日

查看全文
  • Lopamudra Mohanty,
  • Mohit Agarwal,
  • Vivek Mehta

摘要

Deploying trained Artificial Intelligence (AI) models on resource-constrained devices is challenging due to limitations in memory and computational power. To overcome these constraints while maintaining model accuracy, Pruned AI (PAI) models are essential. In response to these challenges, researchers have developed several compression models for Convolutional Neural Networks (CNNs) In this study, a firefly algorithm-based approach was used to minimize the CNN model’s size by choosing a portion of neural network filters and nodes in the hidden layers, keeping the model’s original accuracy for the task of image classification. The ResNet50, VGG16, VGG19 and SqueezeNet architectures were used for model reduction, demonstrating significant gains in reduced storage space without compromising accuracy. The study demonstrates that this approach can achieve an additional 25% reduction in storage space even for SqueezeNet, a model that is already compressed. Moreover, the model’s accuracy improved by 0.23% on the MNIST dataset after pruning with SqueezNet and ResNet50. Similarly, the accuracy also improved for the other pre-trained models like VGG19 1.7% on the MNIST dataset, 7.78% for the CIFAR 10 dataset, and 3.09% for the CIFAR 100 dataset. Finally, an explainable AI model such as Grad-CAM was implemented on uncompressed and compressed models to verify the results and achieve higher confidence.