Choosing an Optimal Neural Network Size to Aid a Search through a Large Image Database
Kieron Messer, Josef Kittler · 1998
In this paper a fast method of selecting a neural network architecture for pattern recognition tasks is presented. We demonstrate that our proposed method of selecting both input features and hidden neurons avoids the pitfalls exhibited by other methods reported in the literature. It is also shown that the resulting network architecture is extremely lean while at the same time significantly improving the network performance. The resulting solution provides a very useful tool which is now being incorporated in the operations system used for large image database surveys. 1 Introduction Pattern recognition neural networks are used to obtain a non-linear function mapping between the input pattern space and the output decision space. In an #-layer feed-forward network (in which the inputs are not considered a layer) there are # # # # ### ## ### #### # free parameters (weights and biases) that try to model this mapping, where # # is the number of units in layer #. For example, a ...