Expert visual classification and neural networks: can general solutions be found?
R. Ellis, Robert Simpson, Phil F. Culverhouse, Thomas Parisini, R. Williams, Beatriz Reguera, Barbara Moore, David G. Lowe · 2002
The authors discuss the potential of artificial neural networks for automating expert visual classifications undertaken in the marine sciences. They illustrate their application with two examples and show their performance, in those cases, to be strongly constrained by the nature of the training data. A proposal for escaping these data limitations is described which involves providing networks with multiple, coarse input channels. They provide evidence that performance is enhanced by this technique and suggest that it may be possible to construct networks which have a general ability to learn specific visual discriminations.>