Stabilizing Invertible Neural Networks Using Mixture Models

In this paper, we analyze the properties of invertible neural networks, which provide a way solving inverse problems. Our main focus lies on investigating and controlling Lipschitz constants corresponding networks. Without such an control, numerical simulations are prone to errors not much is gained against traditional approaches. Fortunately, our analysis indicates that changing latent distribution from standard normal one Gaussian mixture model resolves issue exploding constants. Indeed, confirm modification leads significantly improved sampling quality in multimodal applications.