Enhancing Accuracy for Classification Using the CNN Model and Hyperparameter Optimization Algorithm

Abstract

The Convolutional Neural Network (CNN) is a widely employed deep learning model, particularly effective for image recognition and classification tasks. The performance of a CNN is influenced not only by its architecture but also critically by its hyperparameters. Consequently, optimizing hyperparameters is essential for improving CNN model performance. In this study, the authors propose leveraging optimization algorithms such as Random Search, Bayesian Optimization with Gaussian Processes, and Bayesian Optimization with Treestructured Parzen Estimators to fine-tune the hyperparameters of the CNN model. The performance of the optimized CNN is compared with traditional machine learning models, including Random Forest (RF), Support Vector Classification (SVC), and K-Nearest Neighbors (KNN). Both the MNIST and Olivetti Faces datasets are utilized in this research. In the training procedure, on the MNIST dataset, the CNN model achieved a minimum accuracy of 97.85%, surpassing traditional models, which had a maximum accuracy of 97.50% across all optimization techniques. Similarly, on the Olivetti Faces dataset, the CNN achieved a minimum accuracy of 94.96%, while traditional models achieved a maximum accuracy of 94.00%. In the training-testing procedure, the CNN demonstrated impressive results, achieving accuracy rates exceeding 99.31% on the MNIST dataset and over 98.63% on the Olivetti Faces dataset, significantly outperforming traditional models, whose maximum values were 98.69% and 97.50%, respectively. Furthermore, the study compares the performance of the CNN model with three optimization algorithms. The results show that integrating CNN with these optimization techniques significantly improves prediction accuracy compared to traditional models.