Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations
Authors: Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Schölkopf, Olivier Bachem
ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Then, we train more than 12 000 models covering most prominent methods and evaluation metrics in a reproducible large-scale experimental study on seven different data sets. |
| Researcher Affiliation | Collaboration | 1ETH Zurich, Department for Computer Science 2Max Planck Institute for Intelligent Systems 3Google Research, Brain Team. |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | We release disentanglement_lib2, a new library to train and evaluate disentangled representations. https://github.com/google-research/ disentanglement_lib |
| Open Datasets | Yes | We consider four data sets in which x is obtained as a deterministic function of z: d Sprites (Higgins et al., 2017a), Cars3D (Reed et al., 2015), Small NORB (Le Cun et al., 2004), Shapes3D (Kim & Mnih, 2018). |
| Dataset Splits | No | The paper mentions training 'more than 12 000 models on seven data sets' and discusses 'statistical efficiency' for downstream tasks using 100 vs 10,000 samples. However, it does not provide explicit training, validation, or test dataset split percentages or counts for its primary experiments. |
| Hardware Specification | Yes | Reproducing these experiments requires approximately 2.52 GPU years (NVIDIA P100). |
| Software Dependencies | No | The paper mentions software components like 'convolutional architecture' and 'optimizer' but does not provide specific version numbers for any software dependencies, libraries, or programming languages. |
| Experiment Setup | No | The paper mentions general setup details like 'Each method uses the same convolutional architecture, optimizer, hyperparameters of the optimizer and batch size. All methods use a Gaussian encoder where the mean and the log variance of each latent factor is parametrized by the deep neural network, a Bernoulli decoder and latent dimension fixed to 10. We choose six different regularization strengths, i.e., hyperparameter values, for each of the considered methods.' However, it does not provide specific concrete values for the learning rate, batch size, or the specific regularization strengths used. It defers 'full details' to Appendix G, which is not provided. |