muARTMAP: use of mutual information for category reduction in Fuzzy ARTMAP
E Gomez-Sanchez1, Y A Dimitriadis, J M Cano-Izquierdo
1Dept. of Signal Theory, Commun. and Telematics Eng., Valladolid Univ.
IEEE Transactions on Neural Networks
|February 5, 2008
Summary
A novel muARTMAP architecture addresses category proliferation in Fuzzy ARTMAP by optimizing mutual information and allowing some error to prevent overfitting. This approach enhances robustness and efficiency, particularly in high-dimensional data and real-world applications like character recognition.
Area of Science:
- Computational Intelligence
- Machine Learning
- Pattern Recognition
Background:
- Fuzzy ARTMAP (FAM) suffers from category proliferation, leading to inefficient models.
- Overfitting is a common issue in supervised learning, especially with complex datasets.
Purpose of the Study:
- Introduce muARTMAP, a new architecture to overcome FAM's limitations.
- Optimize input space partitioning for improved mutual information with the output space.
- Enhance model robustness and efficiency in high-dimensional and noisy environments.
Main Methods:
- Probabilistic framework to partition input space.
- Optimization of mutual information between input and output spaces.
- Implementation of an inter-ART reset mechanism for exception handling.
Main Results:
- muARTMAP demonstrates superior performance over FAM and Boosted ARTMAP on synthetic benchmarks.
- Exhibits increased robustness to noise and better scalability with increasing dimensionality.
- Achieves comparable performance to FAM in handwritten character recognition with a more compact rule set.
Conclusions:
- muARTMAP effectively addresses category proliferation and overfitting in FAM.
- The architecture offers improved robustness, efficiency, and scalability.
- Presents a viable alternative for real-world pattern recognition tasks requiring compact models.
