eduzhai > Applied Sciences > Engineering >

Simple and Effective VAE Training with Calibrated Decoders

  • Save

... pages left unread,continue reading

Document pages: 16 pages

Abstract: Variational autoencoders (VAEs) provide an effective and simple method formodeling complex distributions. However, training VAEs often requiresconsiderable hyperparameter tuning, and often utilizes a heuristic weight onthe prior KL-divergence term. In this work, we study how the performance ofVAEs can be improved while not requiring the use of this heuristichyperparameter, by learning calibrated decoders that accurately model thedecoding distribution. While in some sense it may seem obvious that calibrateddecoders should perform better than uncalibrated decoders, much of the recentliterature that employs VAEs uses uncalibrated Gaussian decoders with constantvariance. We observe empirically that the naïve way of learning variance inGaussian decoders does not lead to good results. However, other calibrateddecoders, such as discrete decoders or learning shared variance cansubstantially improve performance. To further improve results, we propose asimple but novel modification to the commonly used Gaussian decoder, whichrepresents the prediction variance non-parametrically. We observe empiricallythat using the heuristic weight hyperparameter is not necessary with ourmethod. We analyze the performance of various discrete and continuous decoderson a range of datasets and several single-image and sequential VAE models.Project website: this https URL

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×