Variational Hyper-encoding Networks

We propose a framework called HyperVAE for encoding distributions of distributions. When target distribution is modeled by VAE, its neural network parameters are sampled from in the model space hyper-level VAE. variational inference to implicitly encode parameter into low dimensional Gaussian distribution. Given distribution, we predict posterior latent code, then use matrix-network decoder generate parameters. can full contrast common hyper-networks practices, which only scale and bias vectors modify target-network Thus preserves information about each task space. derive training objective using minimum description length (MDL) principle reduce complexity HyperVAE. evaluate density estimation tasks, outlier detection discovery novel design classes, demonstrating efficacy.