The Earth Mover’s Pinball Loss: Quantiles for Histogram-Valued Regression
Authors: Florian List
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our method with an illustrative toy example, a football-related task, and an astrophysical computer vision problem. We show that with our loss function, the accuracy of the predicted median histograms is very similar to the standard EMD case (and higher than for perbin loss functions such as cross-entropy), while the predictions become much more informative at almost no additional computational cost. |
| Researcher Affiliation | Academia | 1The University of Sydney, Sydney Institute for Astronomy, School of Physics, A28, NSW 2006, Australia. |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | No | The paper lists software used like Tensorflow and Keras, and mentions 'The EMPL is easy to implement (see the Supplementary Material)', but does not provide a direct link to the source code for their proposed method or explicitly state that their code is open-source/publicly available. |
| Open Datasets | Yes | We use data from all the Bundesliga seasons between 1995 96 (when the 3-points-for-a-win rule was introduced) and 2017 18, keeping the seasons 1998 99, 2006 07, and 2014 15 for testing, while using the other 20 seasons as training data.2 Data: www.kaggle.com/thefc17/bundesliga-results-19932018. |
| Dataset Splits | No | The paper specifies training and testing data splits (e.g., '1/15th of the remaining maps for testing and use the others as training data', 'keeping the seasons ... for testing, while using the other 20 seasons as training data') but does not explicitly mention a separate validation set split. |
| Hardware Specification | No | The author acknowledges the National Computational Infrastructure (NCI), which is supported by the Australian Government, for providing services and computational resources on the supercomputer Gadi that have contributed to the research results reported within this paper. This mentions a supercomputer by name ('Gadi') but does not provide specific hardware details like GPU/CPU models or memory. |
| Software Dependencies | No | Software: matplotlib (Hunter, 2007), seaborn (Waskom et al., 2017), numpy (Oliphant, 2006), scipy (Virtanen et al., 2020), numba (Lam et al., 2015), healpy (Zonca et al., 2019), Tensorflow (Abadi et al., 2016), Keras (Chollet et al., 2015), ray (Moritz et al., 2017), dill (Mc Kerns et al., 2011), cloudpickle, colorcet. This list names software but does not include specific version numbers for them (e.g., 'Tensorflow 2.x' or 'numpy 1.20'). The years in parentheses refer to publication dates of the references, not software versions. |
| Experiment Setup | Yes | We train a simple multilayer perceptron (MLP) containing 2 hidden layers with 128 neurons each, Re LU activation and batch normalization for the hidden layers, and a softmax activation for the output layer... The NN training consists of 10,000 batch iterations at a batch size of 2,048. We minimize the EMPL with randomly drawn quantile levels τ using an Adam optimizer... We minimize the Smoothed EMPL with α = 0.005 over 250 epochs using a batch size of 2,048. |