Ok hear me out. Before you start explaining how the feature works like I should know better: I know how they work, you add the activations of the previous layer to the actions of a later layer to stabilize information/gradient flow. My Question is, why doesn't it just get implemented in the seemingly more sensible way of CONCATENATING the some previous layer's activations onto a later layer's activations? This avoid the strange choice of addition which has questionable benefit, and it is the more straight forward and obvious way to make the previous layer's activations available. Of course it does create more parameters to train, but that seems like a small sacrifice to me.