Cooperative information maximization with Gaussian activation functions for self-organizing maps.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Information Theory

Background:

• Self-organizing maps (SOMs) are unsupervised learning algorithms used for dimensionality reduction and data visualization.
• Traditional SOMs often lack explicit mechanisms for information maximization and cooperative learning.
• Enhancing information content and clarity in SOMs is crucial for effective data analysis.

Purpose of the Study:

• To propose a novel information-theoretic method for generating explicit self-organizing maps (SOMs).
• To enhance the information content and clarity of SOMs through cooperative and entropy maximization principles.
• To evaluate the proposed method on diverse datasets, including uniform distribution learning, chemical compound classification, and road classification.

Main Methods:

• Utilizing information theory, specifically mutual information maximization, to drive the competition between input patterns and competitive units.
• Employing a Gaussian function for competitive unit outputs, where proximity to input patterns increases firing rates.
• Incorporating cooperative processes by considering neighboring neuron firing rates to enhance information content.
• Applying entropy maximization when cooperative operations are insufficient to boost information, ensuring more uniform unit usage.

Main Results:

• Experimental results demonstrate that cooperative processes significantly increase the information content within input patterns.
• The method successfully applied to uniform distribution learning, chemical compound classification, and road classification tasks.
• Entropy maximization was shown to effectively increase information and produce clearer SOMs by promoting balanced usage of competitive units.

Conclusions:

• The proposed information-theoretic method offers an effective way to generate explicit self-organizing maps with enhanced information content.
• Cooperative learning and entropy maximization are valuable strategies for improving SOM performance and data representation.
• The method's applicability across various domains highlights its potential for robust data analysis and visualization.