The Variational InfoMax AutoEncoder
This addresses a fundamental limitation in VAEs for researchers and practitioners in generative modeling, though it appears incremental as it builds on existing VAE frameworks.
The paper tackles the issue in Variational AutoEncoders where only one of the inference or generative models can be learned optimally, proposing the Variational InfoMax (VIM) objective to learn a maximal informative generator while bounding network capacity, resulting in an explicit definition of network capacity as an estimation of robustness.
The Variational AutoEncoder (VAE) learns simultaneously an inference and a generative model, but only one of these models can be learned at optimum, this behaviour is associated to the ELBO learning objective, that is optimised by a non-informative generator. In order to solve such an issue, we provide a learning objective, learning a maximal informative generator while maintaining bounded the network capacity: the Variational InfoMax (VIM). The contribution of the VIM derivation is twofold: an objective learning both an optimal inference and generative model and the explicit definition of the network capacity, an estimation of the network robustness.