Duality in RKHSs with Infinite Dimensional Outputs: Application to Robust Losses

Authors: Pierre Laforgue, Alex Lambert, Luc Brogat-Motte, Florence D’Alché-Buc

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Robustness benefits are emphasized by a theoretical stability analysis, as well as empirical improvements on structured data applications. In this section, we discuss some applications unlocked by vv-RKHSs with infinite dimensional outputs. In particular, structured prediction, structured representation learning, and functional regression are formally described, and numerical experiments highlight the benefits of the losses introduced.
Researcher Affiliation Academia 1LTCI, T el ecom Paris, Institut Polytechnique de Paris, France.
Pseudocode Yes Algorithm 1 Projected Gradient Descents (PGDs)
Open Source Code No The paper does not contain an explicit statement about releasing source code for the described methodology or a link to a code repository.
Open Datasets Yes YEAST dataset Finley and Joachims (2008); Metabolite dataset (Schymanski et al., 2017); drug dataset, introduced in Su et al. (2010); EMG dataset (Ramsay and Silverman, 2007).
Dataset Splits Yes Hyperparameters Λ, ϵ, κ have been selected among geometrical grids by cross-validation on the train dataset solely, and performances evaluated on the same test set as the above publications. Usefulness of minimizing the Huber loss is illustrated in Figure 8 by computing the Leave-One-Out (LOO) error associated to each model for various values of m.
Hardware Specification No The paper does not provide any specific details about the hardware used for running experiments, such as CPU or GPU models, or cloud computing instance types.
Software Dependencies No The paper mentions "Pystruct lib implementation (M uller and Behnke, 2014)" but does not provide a specific version number for this or any other software dependency relevant to reproducibility.
Experiment Setup No The paper states that "Hyperparameters Λ, ϵ, κ have been selected among geometrical grids by cross-validation on the train dataset solely" and that specific Λ values were "picked for their interesting behavior" with ϵ and κ "chosen to produce the best scores." However, it does not provide the concrete numerical values of these hyperparameters (e.g., learning rate, batch size, or the chosen Λ, ϵ, κ values for the reported results) within the main text.