Classification of Sonar Data

The sonar data set ([sejn96]) is a collection of sonar signals, coded as 60 continuous attributes. The task is to discriminate between signals bounced off a rock or a metal cylinder (2 classes). The test was done with a number of networks and very similar results obtained independent of the architecture. Table 7.3 shows the number of neurons in the network, the number of weights to train, and the performance achieved on the test set.

 

Number of Number of Number of Achieved

Network Neurons weights Result

Modular: 15 + 1 74 340 87.5%

Modular: 10 + 1 74 440 88.5%

Modular: 4 + 1 62 664 88.5%

Modular: 3 + 1 48 620 90.4%

Single: 1 23 950 90.4%