Convexity in ReLU Neural Networks: beyond ICNNs?