Deep Neural Networks as Point Estimates for Deep Gaussian Processes