We introduce marginalizing ows, an extension to normalizing ows which allow for better density estimation by marginaliziation of auxiliary random variables. We give an outline of the shortcomings of normalizing ows and motivate our approach. We trained models with an architecture based on Real NVP by Dinh et al. on several datasets. For low-dimensional data we see that marginalizing ows consistently predict higher likelihood than normalizing ows, but our results do not generalize to higher-dimensional data like images.
Faculteit der Sociale Wetenschappen