Pruning nodes in feedforward neural networks (Q2763604)

Method of induced pruning in feedforward neural networks, as an alternative to the zero-pruning algorithm, is proposed. The induced pruning is realized via the following algorithm: 1) The network is trained until the error function \(E\) reaches a predetermined value; 2) The training set is modified in such a way that the \(i\)-th coordinates of all input patterns get inhibitory activation, while output patterns remain unchanged; 3) The network is retrained with the modified training set; 4) The \(i\)-th input node is pruned. An advantage of this approach is that there is no need for intervention in the learning algorithm, i.e. the weights are manipulated ``from outside''. Also, the pruned nodes, which were not actually removed, can be reconnected and the network retrained with the initial training set, if needed. The performance degradation rate of the trained network, when a number of input nodes are pruned, is also investigated.NEWLINENEWLINEFor the entire collection see [Zbl 0977.00022].