Towards Nonlinear Disentanglement in Natural Data with Temporal Sparse Coding
Authors: David A. Klindt, Lukas Schott, Yash Sharma, Ivan Ustyuzhaninov, Wieland Brendel, Matthias Bethge, Dylan Paiton
ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show that equipping practical estimation methods with our prior often surpasses the current state-of-the-art on several established benchmark datasets without any impractical assumptions, such as knowledge of the number of changing generative factors. Furthermore, we contribute two new benchmarks, Natural Sprites and KITTI Masks, which integrate the measured natural dynamics to enable disentanglement evaluation with more realistic datasets. We test our theory on these benchmarks and demonstrate improved performance. |
| Researcher Affiliation | Academia | David Klindt University of Tübingen klindt.david@gmail.com Lukas Schott University of Tübingen lukas.schott@bethgelab.org Yash Sharma University of Tübingen yash.sharma@bethgelab.org Ivan Ustyuzhaninov University of Tübingen ivan.ustyuzhaninov@bethgelab.org Wieland Brendel University of Tübingen wieland.brendel@bethgelab.org Matthias Bethge University of Tübingen matthias.bethge@bethgelab.org Dylan M Paiton University of Tübingen dylan.paiton@bethgelab.org |
| Pseudocode | No | No pseudocode or algorithm blocks are explicitly labeled or formatted as such in the paper. |
| Open Source Code | Yes | Code: https://github.com/bethgelab/slow_disentanglement |
| Open Datasets | Yes | Most of the standard datasets for disentanglement (d Sprites (Matthey et al., 2017), Cars3D (Reed et al., 2015), Small NORB (Le Cun et al., 2004), Shapes3D (Kim and Mnih, 2018), MPI3D (Gondal et al., 2019)) have been compiled into a disentanglement library (Dis Lib) by Locatello et al. (2018). ...Natural Sprites: The benchmark is available at https://zenodo.org/record/3948069. ...KITTI Masks: The benchmark is available at https://zenodo.org/record/3931823. |
| Dataset Splits | No | No specific train/validation/test dataset splits (e.g., by percentage or absolute count) are explicitly provided in the paper. The paper mentions 'held-out test data' in the context of evaluation metrics, but not for the overall dataset partitioning for training and validation. |
| Hardware Specification | No | No specific hardware details (such as GPU or CPU models, or memory specifications) are provided for the experimental setup. |
| Software Dependencies | No | All models are implemented in Py Torch (Paszke et al., 2019). ... which is implemented via the scikit-learn Python package (Pedregosa et al., 2011) in the Disentanglement Library (Dis Lib, Locatello et al., 2018). (Specific version numbers for PyTorch or scikit-learn are not given, only the year of publication for the respective papers). |
| Experiment Setup | Yes | All models are implemented in Py Torch (Paszke et al., 2019). To facilitate comparison, the training parameters, e.g. optimizer, batch size, number of training steps, as well as the VAE encoder and decoder architecture are identical to those reported in (Locatello et al., 2018; 2020). We use this architecture for all datasets, only adjusting the number of input channels (greyscale for d Sprites, small NORB, and KITTI Masks; three color channels for all other datasets). ... In order to select the conditional prior regularization and the prior rate in an unsupervised manner, we perform a random search over γ [1, 16] and λ [1, 10] and compute the recently proposed unsupervised disentanglement ranking (UDR) scores (Duan et al., 2020). We notice that the optimal values are close to γ = 10 and λ = 6 on most datasets, and thus use these values for all experiments. |