Learning Wasserstein Embeddings
Authors: Nicolas Courty, Rémi Flamary, Mélanie Ducoffe
ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Numerical experiments supporting this idea are conducted on image datasets, and show the wide potential benefits of our method. |
| Researcher Affiliation | Academia | Nicolas Courty Université de Bretagne Sud IRISA, UMR 6074, CNRS ncourty@irisa.fr; Rémi Flamary Université Côte d Azur, OCA Lagrange, UMR 7293, CNRS remi.flamary@unice.fr; Mélanie Ducoffe Université Côte d Azur I3S, UMR 7271, CNRS melanie.ducoffe@i3s.unice.fr |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions using the 'POT toolbox (Flamary & Courty, 2017)' but does not state that the code for the described methodology (DWE) is open-source or provide a link to it. |
| Open Datasets | Yes | Our first numerical experiment is performed on the well known MNIST digits dataset. |
| Dataset Splits | Yes | Among this million, 700 000 are used for learning the neural network, 200 000 are used for validation and 100 000 pairs are used for testing purposes. |
| Hardware Specification | Yes | The DWE model is learnt on a GTX Titan X Maxwell 980 GPU node and takes around 1h20 with a stopping criterion computed from on a validation set. |
| Software Dependencies | No | The paper mentions using the 'POT toolbox (Flamary & Courty, 2017)' but does not provide specific version numbers for this or any other software dependencies. |
| Experiment Setup | Yes | The architecture used for the embedding φ consists in 2 convolutional layers with Re LU activations: first a convolutional layer of 20 filters with a kernel of size 3 by 3, then a convolutional layer of 5 filters of size 5 by 5. The convolutional layers are followed by two linear dense layers respectively of size 100 and the final layer of size p = 50. The architecture for the reconstruction ψ consists in a dense layer of output 100 with Re LU activation, followed by a dense layer of output 5*784. We reshape the layer to map the input of a convolutional layer: the output vector is (5,28,28) 3D-tensor. Eventually, we invert the convolutional layers of φ with two convolutional layers: first a convolutional layer of 20 filters with Re LU activation and a kernel of size 5 by 5, followed by a second layer with 1 filter, with a kernel of size 3 by 3. Eventually the decoder outputs a reconstruction image of shape 28 by 28. In this work, we only consider grayscale images, that are normalized to represent probability distributions. Hence each image is depicted as an histogram. In order to normalize the decoder reconstruction we use a softmax activation for the last layer. |