Regularization properties of adversarially-trained linear regression

Authors: Antonio Ribeiro, Dave Zachariah, Francis Bach, Thomas Schön

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

Reproducibility Variable Result LLM Response
Research Type Experimental We confirm our theoretical findings with numerical examples. (Abstract) and Numerical Experiments (Section Title)
Researcher Affiliation Academia Antônio H. Ribeiro Uppsala University antonio.horta.ribeiro@it.uu.se Dave Zachariah Uppsala University dave.zachariah@it.uu.se Francis Bach PSL Research University, INRIA francis.bach@inria.fr Thomas B. Schön Uppsala University thomas.schon@it.uu.se
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Code for reproducing the figures is available in: https://github.com/antonior92/advtrain-linreg
Open Datasets Yes Regularization paths estimated in the Diabetes dataset [18]. (Figure 1 caption). We illustrate our method on the Diverse MAGIC wheat dataset [44] from the National Institute for Applied Botany. (Section 7).
Dataset Splits No For Lasso, ridge and adversarial training, we use the best δ or λ available for each method (obtained via grid search). We use a random ξ, since we do not know the true additive noise. Even with this approximation, ℓ -adversarial training performs comparably with Lasso with the regularization parameter set using 5-fold cross-validation doing a full search in the hyperparameter space. While 5-fold cross-validation is mentioned for hyperparameter tuning, there is no explicit description of how the main datasets (e.g., Diabetes, MAGIC wheat) were split into train/validation/test sets with percentages or sample counts.
Hardware Specification No The paper does not provide any specific hardware details used for running its experiments.
Software Dependencies No In all the numerical examples the adversarial training solution is implemented by minimizing (2) using CVXPY [42]. The paper mentions CVXPY but does not specify its version number or any other software dependencies with version numbers.
Experiment Setup Yes For Lasso, ridge and adversarial training, we use the best δ or λ available for each method (obtained via grid search). (Section 7). We generate the data synthetically using an isotropic Gaussian feature model (see Section 7) with n = 60 training data points and p = 200 features. (Figure 4 caption).