Related Experiment Videos
Efficient training of RBF neural networks for pattern recognition.
1Istituto di Analisi dei Sistemi ed Informatica, CNR, 00185 Rome, Italy. lampariello@iasi.rm.cnr.it
IEEE Transactions on Neural Networks
|February 6, 2008
Summary
This study introduces a novel training method for Radial Basis Function (RBF) neural networks in pattern recognition. The new approach improves computational efficiency for distinguishing data sets.
Area of Science:
- Computer Science
- Artificial Intelligence
- Machine Learning
Background:
- Training Radial Basis Function (RBF) neural networks is crucial for pattern recognition tasks.
- Classical methods often use least-squares error functions, which may not be optimal for classification.
- Classification requires network outputs to exceed or fall below a specific threshold.
Purpose of the Study:
- To propose an alternative training approach for RBF neural networks in classification.
- To develop a method that addresses the specific needs of classification problems.
- To improve the efficiency of RBF network training.
Main Methods:
- Formulating the RBF training problem as a system of nonlinear inequalities.
- Defining a novel error function that focuses on violated inequalities.
- Developing a training algorithm based on this inequality formulation.
- Comparing the proposed algorithm with the traditional least-squares method.
Main Results:
- The proposed approach demonstrates effectiveness in RBF network training for pattern recognition.
- The new method shows significant savings in computational time compared to least-squares.
- The algorithm successfully trains RBF networks for distinguishing between two disjoint sets.
Conclusions:
- The proposed inequality-based training method is effective for RBF neural networks in classification.
- This approach offers a computationally efficient alternative to least-squares methods.
- The findings highlight the potential for improved RBF network training algorithms.