Discrete variational auto-encoders (VAEs) are able to represent semantic latent spaces in generative learning. In many real-life settings, the discrete latent space consists of high-dimensional structures, and propagating gradients through the relevant structures often requires enumerating over an exponentially large latent space. Recently, various approaches were devised to propagate approximated gradients without enumerating over the space of possible structures. In this work, we use Natural Evolution Strategies (NES), a class of gradient-free black-box optimization algorithms, to learn discrete structured VAEs. The NES algorithms are computationally appealing as they estimate gradients with forward pass evaluations only, thus they do not require to propagate gradients through their discrete structures. We demonstrate empirically that optimizing discrete structured VAEs using NES is as effective as gradient-based approximations. Lastly, we prove NES converges for non-Lipschitz functions as appear in discrete structured VAEs.


    Access

    Download


    Export, share and cite



    Title :

    Learning Discrete Structured Variational Auto-Encoder using Natural Evolution Strategies


    Contributors:
    Berliner, Alon (author) / Rotman, Guy (author) / Adi, Yossi (author) / Reichart, Roi (author) / Hazan, Tamir (author)

    Publication date :

    2022


    Remarks:

    Published as a conference paper at ICLR 2022


    Type of media :

    Preprint


    Type of material :

    Electronic Resource


    Language :

    English




    AVAE: Adversarial Variational Auto Encoder

    Plumerault, Antoine / Borgne, Hervé Le / Hudelot, Céline | ArXiv | 2020

    Free access

    Disentangled Variational Auto-Encoder for Semi-supervised Learning

    Li, Yang / Pan, Quan / Wang, Suhang et al. | ArXiv | 2017

    Free access

    Variational auto encoder for mixed data types

    European Patent Office | 2020

    Free access

    Disentangling Latent Factors of Variational Auto-Encoder with Whitening

    Hahn, Sangchul / Choi, Heeyoul | ArXiv | 2018

    Free access