Artificial neural networks with input gates

An architecture of multi-layer neural networks is proposed. The networks are equipped with gates on their input channels in order to control the flow of input signals. A gate on an input channel opens and closes depending on the current values of the other input signals. The dependency is automatically determined based on training data. These gates give the networks a good generalization ability because they can eliminate harmful inputs. Also they can indicate which input is significant in which situations, and therefore they provide us with an insight into the input-output relationship underlying the training data.