Exploring The Internal Dynamics of Self-Organizing Maps for Image Clustering

The increasing reliance on image classification networks across diverse fields necessitates well-labeled data for effective training. However, high-quality labeling remains a labor-intensive, human-dependent process. This study presents a methodological examination of Self-Organizing Maps (SOMs), an unsupervised learning technique, to understand its limitations and capabilities in image clustering. We constructed a synthetic dataset of clearly classified images with varying noise levels to systematically evaluate SOM performance. We developed novel metrics to examine the network's internal dynamics, focusing on convergence trends and sensitivity. The SOM's clustering performance was assessed in terms of cluster quality and convergence speed, with results compared to K-means clustering. Our investigation delved into the SOM's intrinsic characteristics, including the degree of change in Best Matching Units (BMUs), the percolation of these changes through the network architecture, and the volatility of BMU connections. This comprehensive analysis aims to provide insights into the SOM's suitability for unsupervised learning in the image domain. Additionally, we applied the SOM to the UTKFace benchmark dataset to evaluate its ability to categorize images similarly to human perception (by race, age, and gender). Our findings indicate that utilizing SOMs for image tagging requires a thorough understanding of the network's behavior and convergence process for specific datasets. This study contributes to the broader understanding of unsupervised learning techniques in image clustering, offering valuable insights for researchers and practitioners in the field of computer vision.