VFlow: More Expressive Generative Flows with Variational Data Augmentation
This addresses a bottleneck in generative modeling for researchers and practitioners by improving flow-based models, though it is an incremental advancement building on existing flows.
The paper tackles the limited expressiveness of generative flows due to architectural constraints by introducing VFlow, which uses variational data augmentation to allow more flexible intermediate representations. It achieves a state-of-the-art result of 2.98 bits per dimension on CIFAR-10 and offers more compact models for similar quality.
Generative flows are promising tractable models for density modeling that define probabilistic distributions with invertible transformations. However, tractability imposes architectural constraints on generative flows, making them less expressive than other types of generative models. In this work, we study a previously overlooked constraint that all the intermediate representations must have the same dimensionality with the original data due to invertibility, limiting the width of the network. We tackle this constraint by augmenting the data with some extra dimensions and jointly learning a generative flow for augmented data as well as the distribution of augmented dimensions under a variational inference framework. Our approach, VFlow, is a generalization of generative flows and therefore always performs better. Combining with existing generative flows, VFlow achieves a new state-of-the-art 2.98 bits per dimension on the CIFAR-10 dataset and is more compact than previous models to reach similar modeling quality.