Pruning via number of links and neuron activities

A new pruning concept that is based on the number of connections or links and the neuron activities and/or sensitivities is proposed. The concept can be applied to a number of existing algorithms which would then generate a consistent subnetwork even when starting with different initial topologies. The concept can easily be implemented using existing neural network chips by fixing the number of input and output links of each neuron. This new concept is presented and its adaptation on some existing algorithms is outlined. Some of the experimental results are included.<
>