The neural tree classifier: Complex decision boundaries with single-layer nets
✍ Scribed by Roberto Molinari; Amro El-Jaroudi
- Publisher
- Elsevier Science
- Year
- 1993
- Tongue
- English
- Weight
- 475 KB
- Volume
- 6
- Category
- Article
- ISSN
- 0952-1976
No coin nor oath required. For personal study only.
✦ Synopsis
Multi-layer neural networks have the ability to create non-linear classification boundaries when used for pattern recognition. These nets suffer from slow training times as well as high computational requirements. Single-layer nets (SLNs) avoid these disadvantages but cannot provide non-linear decision boundaries. This paper presents an approach that uses a series of SLNs arranged in a binary tree structure to approximate the performance of multi-layer nets. The SLNs are trained with a quasi-Newton training algorithm to minimize the Kullback-Leibler error measure, which guarantees the network to quickly converge to the global minimum. By properly dividing the input space into tree controlled clusters and assigning an SLN to each cluster this approach can handle classification tasks where multi-layer nets have previously been used with the advantages of fast training and simplicity of design. The approach is applied to a pattern-recognition problem, namely speech classification.