Learning Discrete and Continuous Factors of Data via Alternating Disentanglement

Authors: Yeonwoo Jeong, Hyun Oh Song

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that the proposed method clearly disentangles discrete factors and significantly outperforms current disentanglement methods based on the disentanglement score and inference network classification score. The source code is available at https://github.com/snumllab/Disentanglement ICML19. Our quantitative results on 1) d Sprites (Matthey et al., 2017) dataset on the disentanglement evaluation metric by (Kim & Mnih, 2018), and on 2) the inference network classification score on the learned discrete factors show state of the art results outperforming recently proposed disentanglement methods: β VAE, Anchor VAE, Factor VAE, and Joint VAE by a large margin.
Researcher Affiliation Academia 1Department of Computer Science and Engineering, Seoul National University, Seoul, Korea. Correspondence to: Hyun Oh Song <hyunoh@snu.ac.kr>.
Pseudocode Yes Algorithm 1 shows the pseudocode for Cascade VAE.
Open Source Code Yes The source code is available at https://github.com/snumllab/Disentanglement ICML19.
Open Datasets Yes We perform experiments on d Sprites (Matthey et al., 2017), MNIST, Fashion MNIST (Xiao et al., 2017), and Chairs (Aubry et al., 2014) datasets.
Dataset Splits Yes MNIST has 60, 000 images of size 28 × 28. We split 50, 000 as training images and 10, 000 as test images.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments (e.g., GPU model, CPU type, memory).
Software Dependencies No The paper mentions 'Model architecture and training details are provided in supplementary B.', but the provided text does not contain any specific software dependencies with version numbers.
Experiment Setup Yes As discussed in Section 3, we individually control the β term on each continuous variables. Let βj denote the coefficient for a variable j. Each βj s start at the high value βh and gets relieved one at a time to the low value βl. After each r iterations, we relieve one variable j by switching the coefficient from βh to βl. The alternating maximization with discrete variables is enabled after a warm-up time denoted as td. Algorithm 1 shows the pseudocode for Cascade VAE. Input : Data {x(i)}N i=1, Encoder(qφ), Decoder(pθ), βl, βh, r, td, optimizer g Initialize parameters φ, θ. Set βj = βh, j and d(i) = 0, i [N] Set j = 1. The dimension of discrete latent representation S, is fixed to 3 following the experiment protocol in Joint VAE for a fair comparison.