EdVAE: Mitigating codebook collapse with evidential discrete variational autoencoders

Gulcin Baykal, Melih Kandemir, Gozde Unal

Abstract: Highlights•We report evidence of the confirmation bias problem caused by the softmax in dVAE.•We demonstrate that the confirmation bias causes the codebook collapse in dVAE.•We propose an hierarchical extension of dVAE with an evidential view called EdVAE.•We mitigate the codebook collapse problem of dVAEs with EdVAE.•We compare EdVAE with state-of-the-art VQ-VAE based models, and outperform them.