Although variational autoencoders (VAEs) represent a widely influential deep generative model, many aspects of the underlying energy function remain poorly understood. In particular, it is commonly believed that Gaussian encoder/decoder assumptions reduce the effectiveness of VAEs in generating realistic samples. In this regard, we rigorously analyze the VAE objective, differentiating situations where this belief is and is not actually true. We then leverage the corresponding insights to develop a simple VAE enhancement that requires no additional hyperparameters or sensitive tuning. Quantitatively, this proposal produces crisp samples and stable FID scores that are actually competitive with a variety of GAN models, all while retaining desirable attributes of the original VAE architecture. A shorter version of this work will appear in the ICLR 2019 conference proceedings (Dai and Wipf, 2019). The code for our model is available at https://github.com/daib13/ TwoStageVAE.


    Access

    Download


    Export, share and cite



    Title :

    Diagnosing and Enhancing VAE Models


    Contributors:
    Dai, Bin (author) / Wipf, David (author)

    Publication date :

    2019


    Type of media :

    Preprint


    Type of material :

    Electronic Resource


    Language :

    English




    DIAGNOSING DEVICE, DIAGNOSING SYSTEM, DIAGNOSING METHOD AND PROGRAM

    YUGAWA SHINJI | European Patent Office | 2024

    Free access


    BATTERY PACK DIAGNOSING METHOD, CELL DIAGNOSING METHOD, BATTERY PACK DIAGNOSING DEVICE, AND CELL DIAGNOSING DEVICE

    INOUE TAKESHI / HIRASAWA SHIGEKI / ISHIMARU TETSUYA et al. | European Patent Office | 2023

    Free access

    BATTERY PACK DIAGNOSING METHOD, CELL DIAGNOSING METHOD, BATTERY PACK DIAGNOSING DEVICE, AND CELL DIAGNOSING DEVICE

    INOUE TAKESHI / HIRASAWA SHIGEKI / ISHIMARU TETSUYA et al. | European Patent Office | 2024

    Free access

    BATTERY PACK DIAGNOSING METHOD, CELL DIAGNOSING METHOD, BATTERY PACK DIAGNOSING DEVICE, AND CELL DIAGNOSING DEVICE

    INOUE TAKESHI / HIRASAWA SHIGEKI / ISHIMARU TETSUYA et al. | European Patent Office | 2022

    Free access