Performance Analysis of RBF and MLP Networks in Pattern Classification
Both the RBF network and the MLP network can be used as pattern classifiers in situations where there are high-dimensional input spaces.
- PDF / 282,214 Bytes
- 7 Pages / 439.37 x 666.142 pts Page_size
- 30 Downloads / 186 Views
Performance Analysis of RBF and MLP Networks in Pattern Classification
19.1
Introduction
Both the RBF network and the MLP network can be used as pattern classifiers in situations where there are high-dimensional input spaces. Analyzes of what network type would produce better results, as well as, what topology would be more appropriate to achieve such performances, are of great value to choose which network is the most suitable for each type of problem. Pattern recognition done by human beings is derived from several correlations made by sensory organs. Of course, human begins are excellent pattern classifiers, doing such tasks with minimal effort. Similarly, using a learning process, artificial neural networks can assign classes by processing input signals. So, when a new sample is presented, the networks can extract necessary information to perform the classification, since they have a generalization ability, which was acquired through that process of learning from examples. MLP networks, presented in detail in Chap. 5, have one or more intermediate layers, where each neuron is under the action of a nonlinear activation function. The high connectivity level between these neurons is also an inherent characteristic of this neural architecture. With regards to RBF networks, as exposed in details in Chap. 6, they typically have a single intermediate layer, where each neuron uses the radial basis activation function. The output layer, in turn, uses the linear activation function to produce the network answer from the input signals. Typically, hidden layer neurons of an MLP network follow the same model. Hidden layer neurons of an RBF network follow different models and could aim different purposes.
© Springer International Publishing Switzerland 2017 I.N. da Silva et al., Artificial Neural Networks, DOI 10.1007/978-3-319-43162-8_19
259
260
19
Performance Analysis of RBF and MLP Networks …
Table 19.1 Distribution of samples into classes (Wine database)
Classes
Frequency (%)
1 2 3
33.14 39.88 26.27
Table 19.2 Distribution of samples into classes (Wisconsin database)
Classes
Frequency (%)
Benign Malign
65.0 35.0
19.2
Characteristics of the MLP and RBF Networks Under Analysis
The input signals used to accomplish the comparison of performances, between the MLP and RBF networks, are derived from two public and noticeable databases, i.e., the Wine and Breast Cancer Wisconsin databases. The Wine database is identified by 14 numeric attributes, being 13 inputs and one output, which represent three distinct wine classes. The distribution of these classes is given in Table 19.1. Wisconsin database also has numeric attributes, from biopsy analysis in patients, being nine inputs and one output. Samples are classified as Benign and Malign, which frequencies are shown in Table 19.2. It must be pointed out that both databases have long been used in investigations involving comparative analysis between methods for pattern classification.
19.3
Computational Results
The same number of neurons in the intermediate layer wa
Data Loading...
