Adaptive FIR neural model for centroid learning in self-organizing maps

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Self-Organizing Maps (SOMs) are unsupervised learning algorithms used for dimensionality reduction and visualization.
• Traditional SOMs can face challenges with convergence and accurately representing data distributions.

Purpose of the Study:

• To introduce a novel training method for topology preserving maps using a new neuron model.
• To enhance the performance and analytical capabilities of Self-Organizing Maps.

Main Methods:

• A sequential formulation of the Self-Organizing Map (SOM) is proposed, termed FIR-SOM.
• Each neuron is modeled as a Finite Impulse Response (FIR) system with adaptively estimated filter coefficients.
• A distortion measure is minimized during sequential learning to optimize the map.

Main Results:

• The FIR-SOM model computes an ordered set of centroids for static distributions.
• Optimized FIR coefficients consistently approximate a moving average filter, irrespective of input distribution.
• Numerical analysis demonstrates superior convergence properties compared to classic SOM and other unsupervised models.

Conclusions:

• The FIR-SOM provides an effective training method for topology preserving maps.
• The model exhibits robust convergence and adaptive learning capabilities.
• Optimal FIR coefficients are valuable for visualizing cluster densities.