Scholarlink

<p>The major contribution of paper is to design an automated machine learning (AutoML) technique using Genetic algorithm with novel inverted swap crossover for generating and finding best Convolutional Neural Network (CNN) architecture. The proposed method is designed to look for the best convolutional and pooling layers, activation functions, and other architectural decisions in the context of CNN for image classification. For this, benchmark CIFAR-10 dataset is used with unadulterated training and testing sets with ratio of 75–25% for unbiased results. Comparative study also presented between proposed crossover method and conventional crossover methods. Meanwhile, overall results are analyzed in terms of mean fitness per generation graph, parallel coordinate graph, convergence generation, total individual trained. The experimental results of Inverted Swap crossover shown the highest fitness score of 0.9887, faster converging in 17 generations (25% faster than others), and cutting training time to 6 h and 23 min (80% reduction) with only 200 individuals trained. Furthermore, best individual model analyzed in terms of classification matrix, best fitness and training time with resulted fitness score up to 98.87% of best individual, it is better in comparison of conventional neural architectural search methods.</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

Advancing Neural Architecture Search Through an Innovative Genetic Algorithm with Inverted Swap Crossover

  • 期刊论文

  • 发表日期：

    2025年7月18日

查看全文
  • Dhananjay Bisen,
  • Praneet Saurabh,
  • Aditya Dubey,
  • Varshali Jaiswal

摘要

The major contribution of paper is to design an automated machine learning (AutoML) technique using Genetic algorithm with novel inverted swap crossover for generating and finding best Convolutional Neural Network (CNN) architecture. The proposed method is designed to look for the best convolutional and pooling layers, activation functions, and other architectural decisions in the context of CNN for image classification. For this, benchmark CIFAR-10 dataset is used with unadulterated training and testing sets with ratio of 75–25% for unbiased results. Comparative study also presented between proposed crossover method and conventional crossover methods. Meanwhile, overall results are analyzed in terms of mean fitness per generation graph, parallel coordinate graph, convergence generation, total individual trained. The experimental results of Inverted Swap crossover shown the highest fitness score of 0.9887, faster converging in 17 generations (25% faster than others), and cutting training time to 6 h and 23 min (80% reduction) with only 200 individuals trained. Furthermore, best individual model analyzed in terms of classification matrix, best fitness and training time with resulted fitness score up to 98.87% of best individual, it is better in comparison of conventional neural architectural search methods.