Precise Accuracy / Robustness Tradeoffs in Regression: Case of General Norms

Authors: Elvis Dohmatob, Meyer Scetbon

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our findings are empirically confirmed with simple experiments that represent a variety of settings. This work covers feature covariance matrices and attack norms of any nature, extending previous works in this area.
Researcher Affiliation Collaboration 1Meta FAIR 2Microsoft Research. Correspondence to: Elvis Dohmatob <dohmatob@meta.com>.
Pseudocode No No pseudocode or algorithm blocks were explicitly labeled or formatted as such.
Open Source Code Yes See attached Jupyter (Python) notebook.
Open Datasets Yes Applying this to the settings considered in Section 5 of (Cui et al., 2022), namely MNIST and Fashion MNIST with kernel K(x, z) = (1 + 10 3x z)5 (the task is to predict the class label via kernel regression), we see from Table 1 of (Cui et al., 2022) that δ = 1+α(2r 1) = 1+1.3(2(0.13) 1) = 0.038 (0, 1) for MNIST and δ = 1+α(2r 1) = 1+1.2(2(0.15) 1) = 0.16 (0, 1) for Fashion MNIST.
Dataset Splits No The paper does not explicitly mention training, validation, and test splits with specific percentages or sample counts for dataset partitioning.
Hardware Specification No All experiments in our paper were run on a single modern CPU laptop. No specific CPU model or detailed specifications were provided.
Software Dependencies No See attached Jupyter (Python) notebook. While this indicates Python, specific version numbers for Python or any libraries used are not provided.
Experiment Setup Yes For this experiment, we fix the input dimension d = 400, while the covariance matrix Σ and generative model w0 are as in (15) for different values of the sparsity parameter s {10, 20, d = 400}. For different values of sample size n from d to 104, we generate (5 runs) an iid dataset Dn = {(x1, y1), . . . , (xn, yn)} and construct a ordinary least-squares (OLS) estimate bwn for w0. [...] The attack strength r is set as in the second part of Theorem 5.1.