Fast Regularized Discrete Optimal Transport with Group-Sparse Regularizers

Authors: Yasutoshi Ida, Sekitoshi Kanai, Kazuki Adachi, Atsutoshi Kumagai, Yasuhiro Fujiwara

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that our method is up to 8.6 times faster than the original method without degrading accuracy.
Researcher Affiliation Industry Yasutoshi Ida1, Sekitoshi Kanai1, Kazuki Adachi1, Atsutoshi Kumagai1, Yasuhiro Fujiwara2 1NTT Computer and Data Science Laboratories 2NTT Communication Science Laboratories
Pseudocode Yes Algorithm 1 is the pseudocode of our algorithm. Although it applies a solver, such as L-BFGS, to Problem (4), the solver efficiently computes ψ(α + βj1m c) by utilizing the upper bounds, as we described in Lemma 2. Algorithm 2 is the gradient computation.
Open Source Code No The paper does not provide any explicit statement or link indicating that the source code for their method is open-source or publicly available.
Open Datasets Yes We used the digits datasets: USPS (U) (Hull 1994) and MNIST (M) (Lecun et al. 1998) as XS and XT. We used the PIE dataset for the face recognition task (Gross et al. 2008). We used the Caltech-Office dataset for the object recognition task with ten class labels (Griffin, Holub, and Perona 2007; Gong et al. 2012).
Dataset Splits No The paper mentions sampling images (e.g., 'randomly sampled 5,000 images from each dataset'), but it does not specify explicit train/validation/test dataset splits (e.g., percentages, counts, or predefined standard splits) for reproducing the experiments.
Hardware Specification Yes Each experiment was conducted with one CPU core and 264 GB of main memory on a 2.20 GHz Intel Xeon server running Linux.
Software Dependencies No The paper mentions using 'L-BFGS' as a solver, but it does not provide specific version numbers for any software dependencies, libraries, or programming languages used in the implementation.
Experiment Setup Yes We evaluated the processing time and accuracy on combinational settings of the hyperparameters ρ = {0.2, 0.4, 0.6, 0.8} and γ = {103, 102, 101, 100, 10 1, 10 2, 10 3} by following the previous work (Courty et al. 2017; Blondel, Seguy, and Rolet 2018). We set r = 10 in this paper.