Abstract
Artificial neural networks with input gates are proposed which can select relevant input variables to the networks. An input is relevant, if it contains inherent and significant information which is not shared with the other inputs and indispensable in determining the output values. By selecting the relevant inputs, or equivalently by removing irrelevant inputs, one can obtain neural networks with higher generalization abilities and interpretabilities. The proposed network first removes redundant inputs that can be replaced by the other inputs, then its input gates placed at its input channels determine whether each of the remained inputs significantly contribute to the output or not. The gates open and close depending on their trained parameters and current input values to the network. Thus, they can determine the relevant inputs dynam-ically (depending on the current input values) and automatically (based on learning). Numerical examples are provided to demonstrate that the network can successfully select the relevant inputs.
