Dynamic Schedule for Effective On-Line Connection Pruning

摘要

In this letter we investigate a neural network simplification schedule that takes place at the same time as regular weight adjustment and with variable pruning strength. The underlying connection model incorporates an explicit trainable factor modulating the classical synaptic weight. Learning in this context results in a reduced size structure with enhanced generalization ability. The effectiveness of the method is empirically explored in an artificial application and a classical real-world benchmark problem.