Sever: A Robust Meta-Algorithm for Stochastic Optimization
Authors: Ilias Diakonikolas, Gautam Kamath, Daniel Kane, Jerry Li, Jacob Steinhardt, Alistair Stewart
ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We apply SEVER on a drug design dataset and a spam classification dataset, and find that in both cases it has substantially greater robustness than several baselines. |
| Researcher Affiliation | Collaboration | 1Department of Computer Science, University of Southern California, Los Angeles, California, USA 2Simons Institute for the Theory of Computing, Berkeley, California, USA 3Departments of Mathematics and Computer Science and Engineering, University of California, San Diego, California, USA 4Microsoft Research AI, Redmond, Washington, USA 5Department of Statistics, University of California, Berkeley, California, USA 6Web3 Foundation, Zug, Switzerland. |
| Pseudocode | Yes | Algorithm 1 SEVER(f1:n, L, σ) |
| Open Source Code | Yes | Code: https://github.com/hoonose/sever |
| Open Datasets | Yes | The drug discovery dataset was obtained from the Ch EMBL database and was originally curated by Olier et al. (2018);... The spam dataset comes from the Enron corpus Metsis et al. (2006) |
| Dataset Splits | No | The paper describes training and test splits for datasets but does not explicitly mention a validation split. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory specifications) used for running its experiments, only general statements about software libraries. |
| Software Dependencies | Yes | We implemented the latter as a quadratic program, using Gurobi (Gurobi Optimization, Inc., 2016) as a backend solver and YALMIP (L ofberg, 2004) as the modeling language. |
| Experiment Setup | Yes | For all filter methods, we iterate the defense r = 4 times, each time removing the p = ε/2 fraction of points with largest score. For consistency, for each defense and each value of ε we ran the defense 3 times on fresh attack points and display the median of the 3 test errors. |