The problem considered here is the age discrimination of lamellibranch larvae. Patterns of larvae are presetned to a multilayer feedforward neural network. Samples are represented by shape descriptors calculated on the basis of a normalized arc length parametrization of their boundary. After training, the network will classify samples on the basis of their characteristic shapes. In neural network applications one often faces the problem of optimal network size, which is an implicit function of problem complexity and available amount of data for training. This paperpresents some possible solutions to cope with this problems. Results obtained are compared with previous experiments on feedforward networks.
展开▼
