Joint Search of CNN Training Parameters and Training Images Preserving Discriminative Features Required for Generalization Performance of CNN Models Using Interactive Genetic Algorithms

Selecting hyperparameters for training convolutional neural networks (CNNs) and determining the amount of training image data required for reliable generalization remain challenging and often time-consuming tasks, typically relying on manual trial-and-error. Generally, image datasets used to train CNNs often contain redundant images; so, there is a need to remove them and reconstruct a slimmed smart dataset using a reduced portion of the available training data. While genetic algorithms (GAs) have been applied to hyperparameter tuning, less attention has been given to how feature-rich images are systematically extracted from an original full dataset and how the usage rate of training data influences generalization performance. In this study, we propose an interactive GA-based framework that simultaneously optimizes key training parameters and the image usage rate, defined as the proportion of training images used during learning. By applying the interactive GA, for example, an operator can not only change the search range of each parameter for optimization even in GA running but also finish the process as needed while observing the optimization process performed by the GA. The approach is implemented within a MATLAB-based environment, allowing parameters to be adjusted dynamically during the optimization process. The experimental results on four different datasets, including CIFAR-10 and EuroSat, show that the proposed method can achieve a classification performance comparable to that of manually tuned models while using a reduced portion of the available training data. Although this is a dataset-specific observation and not a general minimum data threshold, it suggests that, depending on the dataset, similar accuracy levels can be achieved even with an image utilization rate of approximately 43%, suggesting that not all training samples contribute equally to the model’s performance. These findings demonstrate that the proposed interactive GA-based framework can reduce training-data utilization while maintaining comparable classification performance across multiple datasets.