Encoded pattern classification using constructive learning algorithms based on learning vector quantization.

Area of Science:

• Artificial Intelligence
• Computer Science
• Pattern Recognition

Background:

• Artificial neural networks (ANNs) are widely used for pattern recognition.
• Kohonen-type networks offer unsupervised learning capabilities.
• Efficient training and encoding are crucial for ANN performance.

Purpose of the Study:

• To propose a novel encoding technique for pattern recognition using Kohonen-type ANNs.
• To evaluate four different training techniques for ANNs, including Kohonen's self-organizing network and learning vector quantization (LVQ).
• To introduce two constructive learning algorithms, MLVQ and TLVQ, to enhance network performance and minimize size.

Main Methods:

• A radial grid overlay and conversion to a 2D feature array for training input.
• Training ANNs using Kohonen's self-organizing network, LVQ, MLVQ, and TLVQ algorithms.
• Classifying patterns by comparing input to trained neuron labels based on proximity.

Main Results:

• The encoding strategy effectively clusters distorted versions of the same pattern.
• Kohonen's self-organizing network assigns labels to neurons, enabling pattern classification.
• LVQ, MLVQ, and TLVQ algorithms demonstrate efficiency and robustness in classifying noiseless and noisy patterns.

Conclusions:

• The proposed pattern encoding strategy is effective for Kohonen-type ANNs.
• The developed training techniques, particularly MLVQ and TLVQ, improve recognition accuracy and network efficiency.
• The methods are robust for handling noisy and distorted pattern recognition tasks.