Learning with Fitzpatrick Losses

Authors: Seta Rakotomandimby, Jean-Philippe Chancelier, Michel De Lara, Mathieu Blondel

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the effectiveness of Fitzpatrick losses for label proportion estimation.
Researcher Affiliation Collaboration Seta Rakotomandimby Ecole des Ponts seta.rakotomandimby@enpc.fr Jean-Philippe Chancelier Ecole des Ponts jean-philippe.chancelier@enpc.fr Michel De Lara Ecole des Ponts michel.delara@enpc.fr Mathieu Blondel Google Deep Mind mblondel@google.com
Pseudocode No The paper does not contain any pseudocode or algorithm blocks.
Open Source Code No The code will be opened upon release.
Open Datasets Yes The datasets can be downloaded from http://mulan.sourceforge.net/datasets-mlc.html and https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/.
Dataset Splits Yes In (5), we chose the hyperparameter λ {10 4, 10 3, . . . , 104} against the validation set.
Hardware Specification Yes Experiments were conducted on a Intel Xeon E5-2667 clocked at 3.30GHz with 192 GB of RAM running on Linux.
Software Dependencies No Our implementation relies on the Sci Py [29] and scikit-learn [25] libraries.
Experiment Setup Yes In (5), we chose the hyperparameter λ {10 4, 10 3, . . . , 104} against the validation set. We optimize (5) using the L-BFGS algorithm [20]. The gradient of the Fenchel-Young loss is given in (2), while the gradient of the Fitzpatrick loss is given in Proposition 1, Item 4.