If there is a forth channel of some constant values (ones, for instance) in an RGB image, will a neural network learn to nullify the weights related to this channel? Or will it leave the weights as is and adjust a bias? What about a column of constant values in some tabular data? Will it affect a neural network in some way?