Gaussian ARTMAP: A Neural Network for Fast Incremental Learning of Noisy Multidimensional Maps

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Incremental supervised learning is crucial for adapting to evolving data.
• Existing methods like fuzzy ARTMAP have limitations in noise resistance and representational efficiency.
• Multidimensional analog map learning requires robust and efficient algorithms.

Purpose of the Study:

• Introduce Gaussian ARTMAP, a novel neural network architecture.
• Enhance incremental supervised learning capabilities for analog multidimensional maps.
• Improve upon the performance of fuzzy ARTMAP in terms of efficiency and noise resistance.

Main Methods:

• Developed Gaussian ARTMAP by integrating a Gaussian classifier with an adaptive resonance theory (ART) network.
• Defined the ART choice function using the discriminant function of a Gaussian classifier with separable distributions.
• Defined the ART match function similarly, normalizing distributions to unit height.

Main Results:

• Gaussian ARTMAP demonstrates superior noise resistance compared to fuzzy ARTMAP on benchmark databases.
• The new architecture learns a more efficient internal representation of mappings.
• Simulations show a consistently better trade-off between classification rate and the number of categories.

Conclusions:

• Gaussian ARTMAP offers significant advantages in incremental supervised learning for multidimensional maps.
• The architecture provides enhanced efficiency and robustness to noise.
• Gaussian ARTMAP outperforms existing classifiers, including fuzzy ARTMAP, on benchmark tasks like vowel classification.