Metaheuristics-Based Hyperparameter Tuning for Convolutional Neural Networks

Convolutional neural networks have made remarkable strides in the field of deep learning, achieving outstanding successes. However, to ensure the efficiency and high performance of these networks, it is crucial to optimize their hyperparameters. This paper presents a novel approach that focuses on optimizing hyperparameters for convolutional neural networks. The proposed approach leverages the binary bat algorithm, which is recognized as one of the most efficient algorithms among nature-inspired metaheuristic algorithms. By utilizing this approach, a set of optimal hyperparameters can be obtained, enabling the construction of convolutional neural network models that exhibit superior performance for specific applications. To demonstrate the effectiveness of this approach, the study employs it to determine hyperparameters such as the learning rate of optimizers and the number of filters in each convolutional layer. The objective is to build optimal models for the task of handwritten Chinese character classification. The empirical results obtained demonstrate the remarkable capabilities of the proposed approach. The models generated through this method exhibit higher performance in terms of classification accuracy and convergence ability when compared to the LeNet-5 model, as well as models based on Hyperband, Random Search, and Bayesian Optimization.