Fundamentals of Statistics contains material of various lectures and courses of H. Lohninger on statistics, data analysis and here for more.

Growing Neural Networks

Growing neural networks very much resemble the forward selection technique with multiple linear regression. The principal goal of growing neural networks is to perform a feature selection during the growing process.

The method starts with a neural network having only one input neuron. Then each single feature is selected, one after another, and the network is trained and evaluated using this very feature. The one feature which leads to the best results is stored and attached to the first neuron of the input layer after the processing of all features has been completed. After this, the network grows in its input layer by one neuron and the selection process is repeated the same way as described above. Thus, the best features of the previous runs are combined with a new feature which gives the largest increase of model performance. In order to prevent the network from multiple selection of a feature, those features which already have been selected are omitted for the rest of the selection process. The process of the growing network is aborted when a test data set does not show any increase in model performance.

Growing networks are very demanding, as far as computing power is concerned, since each single step of feature selection requires full training of the neural network. In fact, growing networks are not feasible with training techniques such as back propagation. However, they can be computed with fast networks such as RBF nets.