Weights should not be 0 (at least not frequently) but in a ReLU-based neural network, activations are 0 pretty often. You're absolutely right about die area though.