Sinkhorn Barycenter via Functional Gradient Descent
Authors: Zebang Shen, Zhenfu Wang, Alejandro Ribeiro, Hamed Hassani
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct experimental studies to show the efficiency and efficacy of Sinkhorn Descent by comparing with the recently proposed functional Frank-Wolfe method (FW) from [Luise et al., 2019]2. |
| Researcher Affiliation | Academia | Zebang Shen Zhenfu Wang Alejandro Ribeiro Hamed Hassani Department of Electrical and Systems Engineering Department of Mathematics University of Pennsylvania {zebang@seas,zwang423@math,aribeiro@seas,hassani@seas}.upenn.edu |
| Pseudocode | Yes | Algorithm 1 Sinkhorn Descent (SD) Input: measures {βi}n i=1, a discrete initial measure α0, a step size η, and number of iterations S; Output: A measure αS that approximates the Sinkhorn barycenter of {βi}n i=1; for t = 0 to S 1 do αt+1 := T [αt] αt, with T [αt] defined in (11); end for |
| Open Source Code | No | The paper mentions using 'implementation from the Python OT library' and provides a link to it, but it does not provide source code for the methodology described in this paper. |
| Open Datasets | Yes | Additional Visual Results on MNIST We provide additional results on the MNIST dataset. |
| Dataset Splits | No | The paper mentions initializing with distributions and sampling points, but it does not provide specific training, validation, or test dataset splits with percentages or counts. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using the 'Python OT library' and provides a link, but it does not specify any version numbers for this or other software components. |
| Experiment Setup | Yes | In the following, the entropy regularization parameter γ is set to 10^-4 in all tasks to produce results of good visual quality. We run FW for 500 iterations... SD is initialized with a discrete uniform distribution with support size varying from N {20, 40, 80}. |