Class distributions on SOM surfaces for feature extraction and object retrieval.

Area of Science:

• Computational intelligence
• Machine learning
• Information theory

Background:

• Self-Organizing Maps (SOMs) are unsupervised learning algorithms commonly trained on large datasets.
• SOMs can map semantically related data classes by identifying the best matching unit for each data vector.
• The resulting distribution of data vectors on the SOM represents a discrete probability density.

Purpose of the Study:

• To utilize feature distributions on SOMs for comparing different data classes and feature representations.
• To evaluate the effectiveness of these distributions within the PicSOM content-based image retrieval system.
• To apply information-theoretic measures for assessing distribution compactness and independence.

Main Methods:

• Trained SOMs using unsupervised learning on large datasets.
• Mapped subsets of vectors belonging to user-defined classes onto the SOM.
• Employed information-theoretic measures, including entropy and mutual information, to analyze SOM distributions.
• Investigated the impact of low-pass filtering SOM surfaces before entropy calculation.

Main Results:

• Qualitatively different data distributions can be achieved from the same data by varying feature extraction techniques.
• Information-theoretic measures successfully evaluated the compactness of SOM distributions and the independence between distributions.
• Low-pass filtering SOM surfaces influenced the calculation of entropy, providing insights into distribution smoothing.

Conclusions:

• Feature distributions on SOMs offer a valuable method for comparing data classes and feature representations in image retrieval.
• Information-theoretic measures provide robust quantitative tools for analyzing SOM-based data distributions.
• The study demonstrates the utility of SOMs and information theory for enhancing content-based image retrieval systems like PicSOM.