A systematic study of autoencoder hyperparameters for effective feature learning in image recognition tasks: insights from handwriting dataset

Abstract

This research investigates the potential of autoencoders to enhance handwritten digit recognition using the MNIST dataset. Autoencoders, with their encoding and decoding mechanisms, effectively capture essential data patterns, making them powerful tools for feature extraction and dimensionality reduction. The study evaluates various autoencoder architectures, including shallow and deep designs, by fine-tuning hyperparameters such as epochs, batch size, and learning rate to optimize model representations and improve recognition performance. Performance is measured using metrics like Mean Squared Error (MSE) and classification accuracy on reconstructed images, providing a comprehensive assessment of the autoencoder’s effectiveness. Comparative experiments reveal the influence of architectural designs and parameter settings on the model’s ability to generalize and accurately recognize digits. Results demonstrate that autoencoders not only enhance recognition accuracy but also provide valuable insights into the dataset’s structure. These findings underscore the importance of unsupervised neural networks in improving practical applications like image recognition. By optimizing architectural and parameter choices, this study contributes to advancements in digit recognition systems and highlights the broader potential of autoencoders in machine learning.