摘要
It has become an annual tradition for Convolutional Neural Networks(CNNs)to continuously improve their performance in image classification and other applications.These advancements are often attributed to the adoption of more intricate network architectures,such as modules and skip connections,as well as the practice of stacking additional layers to create increasingly complex networks.However,the quest to identify the most optimizedmodel is a daunting task,given that stateof the artConvolutionalNeuralNetwork(CNN)models aremanually engineered.In this research paper,we leveraged a conventional Genetic Algorithm(GA)to craft optimized Convolutional Neural Network(CNN)architectures and pinpoint the ideal set of hyper parameters for image classification tasks using the MNIST dataset.Our experimentation with the MNIST dataset yielded remarkable results.Compared to earlier semi-automatic and automated approaches,our proposed GA demonstrated its efficiency by swiftly identifying the perfect CNN design,accomplishing this feat in just 6 GPU days while achieving an outstanding accuracy of 95.50%.
