DefinePK

Abstract

This research focuses on the parameter optimization of an autoencoder for image classification using a genetic algorithm (GA). An autoencoder is a neural network architecture commonly used for unsupervised learning, dimensionality reduction, and feature extraction. Its performance heavily depends on hyperparameters, which must be carefully tuned to achieve optimal results. In this study, a GA-based optimization approach is proposed to fine-tune the hyperparameters of an autoencoder, including the number of hidden layers, the number of neurons per layer, the activation function, the learning rate, and the batch size. The proposed approach is applied to two datasets: MNIST and EMNIST (an extended version of MNIST for handwritten letters), as well as Fashion-MNIST. The performance is compared against other state-of-the-art optimization techniques. The results demonstrate that GA-based optimization effectively enhances autoencoder performance, outperforming traditional methods in terms of reconstruction error and classification accuracy. Specifically, when using the Adam optimizer, the average classification accuracy achieved is 97.77% with an average computation time of 34.77s, using a learning rate of 0.001, momentum of 0.87, and a sparsity parameter of 0.01. In contrast, the GA-based approach yields an improved accuracy of 98.85% with a slightly higher computation time of 35.44s, using a learning rate of 0.1, momentum of 0.85, and a sparsity parameter of 0.01. This research contributes to the development of an efficient autoencoder optimization framework, applicable to a wide range of tasks, including image classification, data compression, feature extraction, and anomaly detection.