Approximate Geodesics for Deep Generative Models How to efficiently find the shortest path in latent space