Disentanglement by Nonlinear ICA with General Incompressible-flow Networks (GIN)

Authors: Peter Sorrenson, Carsten Rother, Ullrich Köthe

ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on artificial data and EMNIST demonstrate that theoretical predictions are indeed verified in practice.
Researcher Affiliation Academia Peter Sorrenson, Carsten Rother, Ullrich K othe Visual Learning Lab Heidelberg University
Pseudocode No The paper describes network architectures and mathematical formulations but does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any statements or links indicating that the source code for the described methodology is publicly available.
Open Datasets Yes The data comes from the EMNIST Digits training set of 240,000 images of handwritten digits with labels (Cohen et al., 2017).
Dataset Splits No The paper uses the EMNIST Digits training set but does not explicitly specify the proportions or sizes of training, validation, and test splits needed for reproduction.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments.
Software Dependencies No The paper mentions algorithms like Adam optimizer and Real NVP, but does not provide specific software library names with version numbers needed for replication.
Experiment Setup Yes Training converges quickly and stably using the Adam optimizer (Kingma & Ba, 2014) with initial learning rate 10^-2 and other values set to the usual recommendations. Batch size is 1,000 and the data is augmented with Gaussian noise (σ = 0.01) at each iteration. After convergence of the loss, the learning rate is reduced by a factor of 10 and trained again until convergence." and "Optimization is with the Adam optimizer, with initial learning rate 3e-4. Batch size is 240 and the data is augmented with Gaussian noise (σ = 0.01) at each iteration. The model is trained for 45 epochs, then for a further 50 epochs with the learning rate reduced by a factor of 10.