Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Regularized Optimal Transport and the Rot Mover's Distance
Authors: Arnaud Dessein, Nicolas Papadakis, Jean-Luc Rouas
JMLR 2018 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we demonstrate the merits of our methods with experiments using synthetic data to illustrate the effect of different regularizers, penalties and dimensions, as well as real-world data for a pattern recognition application to audio scene classification. In Section 6, we provide experiments to illustrate our methods on synthetic data and real-world audio data in a classification problem. |
| Researcher Affiliation | Collaboration | Arnaud Dessein EMAIL Institut de Math ematiques de Bordeaux CNRS, Universit e de Bordeaux 351 Cours de la Lib eration, 33405 Talence, France Qucit 213 Cours Victor Hugo, 33130 B egles, France |
| Pseudocode | Yes | Algorithm 1 Alternate scaling algorithm. Algorithm 2 Alternate scaling algorithm in the separable case. Algorithm 3 Non-negative alternate scaling algorithm. Algorithm 4 Non-negative alternate scaling algorithm in the separable case. |
| Open Source Code | Yes | 5https://www.math.u-bordeaux.fr/~npapadak/GOTMI/codes.php |
| Open Datasets | Yes | We consider the framework of the DCASE 2016 IEEE AASP challenge with the TUT Acoustic Scenes 2016 database (Mesaros et al., 2016). |
| Dataset Splits | Yes | The audio material is cut into 30-second segments, and is split into two subsets of 75% 25% containing respectively 78 26 segments per class for development and evaluation, resulting in a total of 1170 390 files for training and testing. A 4-fold cross-validation setup is given with the training set. |
| Hardware Specification | No | Nevertheless, for a fair interpretation of the above timing results, we must mention that the two EMD schemes tested were run under MATLAB from native C/C++ implementations1,2 via compiled MEX files3,4. Hence, these EMD codes are quite optimized in comparison to our pure MATLAB prototype codes5 for the RMD. It is thus plausible that optimized C/C++ implementations of our algorithms would be even more competitive in this context. |
| Software Dependencies | No | The baseline system is ran with its default parameters: 40 ms frame size, 20 ms hop size, 60-dimensional MFCCs comprising 20 static (including energy) plus 20 delta and 20 acceleration coefficients extracted with standard settings in RASTAMAT, 16 GMM components learned with standard settings in VOICEBOX. The SVM classifier is implemented with standard settings in LIBSVM, and requires an additional soft-margin parameter C > 0 to be tuned. |
| Experiment Setup | Yes | We use a small tolerance of 10−8 for convergence with the ℓ∞ norm on the marginal difference checked after each iteration as a termination criterion. As a stopping criterion, we use the relative variation with tolerance 10−2 in ℓ2 norm for the main loop of alternate Bregman projections, and the absolute variation with tolerance 10−5 in ℓ2 norm for the auxiliary loops of the Newton Raphson method. The baseline system is ran with its default parameters: 40 ms frame size, 20 ms hop size, 60-dimensional MFCCs comprising 20 static (including energy) plus 20 delta and 20 acceleration coefficients extracted with standard settings in RASTAMAT, 16 GMM components learned with standard settings in VOICEBOX. ... The parameters τ, C ∈ 10{−1,+0,+1,+2} and penalty λ ∈ Λ, where Λ is a manually chosen set of four successive powers of ten depending on the range of the regularizer φ, are tuned automatically by cross-validation. The number of iterations is limited to 100 for the main loop of the algorithm and to 10 for the auxiliary loops of the Newton Raphson method, and the tolerance is set to 10−6 in all loops for convergence with the ℓ∞ norm on the marginal difference checked after each iteration as a termination criterion. |