Sign in Subscribe
Machine learning

Advancing Neural Architecture Search Through Innovative Genetic Algorithm

Strategic Insights

— 1 min read

Researchers from Manipal University in Rajasthan, India, have developed a novel genetic algorithm, called Inverted Swap crossover, for automated machine learning (AutoML) in Convolutional Neural Networks (CNNs). This innovative technique is designed to optimally select the best convolutional and pooling layers, activation functions, and other architectural decisions for image classification tasks. The proposed method outperformed conventional crossover methods, achieving a highest fitness score of 0.9887, faster convergence in 17 generations, and a significant reduction in training time to 6 hours and 23 minutes.

Key Takeaways:

  • The researchers designed an automated machine learning (AutoML) technique using a genetic algorithm with Inverted Swap crossover to optimize CNN architecture for image classification.
  • The proposed method was tested on the CIFAR-10 dataset with unadulterated training and testing sets, achieving an unbiased result with a ratio of 75-25%.
  • Comparative studies were conducted between the proposed Inverted Swap crossover and conventional crossover methods, demonstrating the superiority of the proposed technique.
  • The experimental results showed the Inverted Swap crossover achieving the highest fitness score of 0.9887, faster convergence in 17 generations, and an 80% reduction in training time to 6 hours and 23 minutes.
  • The researchers analyzed the best individual model in terms of classification matrix, best fitness, and training time, achieving a fitness score of up to 98.87%.
  • The proposed method outperformed conventional neural architectural search methods, demonstrating its effectiveness in optimizing CNN architecture for image classification tasks.

Statistics:

  • The proposed Inverted Swap crossover achieved a highest fitness score of 0.9887.
  • The technique demonstrated faster convergence in 17 generations, a 25% improvement compared to other methods.
  • The training time was reduced to 6 hours and 23 minutes, representing an 80% reduction compared to other methods.
  • The researchers achieved a classification matrix with a fitness score of up to 98.87%.
  • The proposed method was compared to conventional crossover methods, demonstrating its superiority.

Sources:

  • Advancing Neural Architecture Search Through an Innovative Genetic Algorithm With Inverted Swap Crossover, National Academy Science Letters, 2025.
  • Springer India, 7TH Floor, Vijaya Building, 17, Barakhamba Road, New Delhi, 110 001, India, (Springer - www.springer.com; National Academy Science Letters - www.springerlink.com/content/0250-541x/)
  • NewsRx. Findings on Mathematics Detailed by Investigators at Manipal University (Advancing Neural Architecture Search Through an Innovative Genetic Algorithm With Inverted Swap Crossover). Life Science Weekly. August 12, 2025; p 1497.

Read more