Unsupervised Learning for MNIST with Exploratory Data Analysis for Digit Recognition

Abstract

This research investigates the application of unsupervised learning techniques for digit recognition using the MNIST dataset. Through a comparative analysis, various dimensionality reduction methods, including ISOmap, PCA, and tSNE, were evaluated for their effectiveness in visualizing and processing the MNIST data. The findings reveal that tSNE consistently outperforms ISOmap and PCA in terms of accuracy, F1- score, precision, and recall, showcasing its superior capability in preserving relevant information within the dataset. Utilizing tSNE for visualizing and clustering digits provides valuable insights into the underlying structure of the dataset, uncovering complex patterns in digit relationships. These results contribute to the advancement of digit recognition systems, offering potential improvements in classification accuracy and model reliability. The success of tSNE highlights the importance of nonlinear dimensionality reduction techniques in handling complex datasets, such as MNIST. This research underscores the significance of unsupervised learning approaches, particularly tSNE, in enhancing digit recognition systems' performance, with implications extending across various application domains. Continued research is recommended to explore further applications and potentials of unsupervised learning techniques and to deepen our understanding of the MNIST dataset's structure and complexity.