Towards Lower Bounds on the Depth of ReLU Neural Networks