Stochastic Bias-Reduced Gradient Methods

Authors: Hilal Asi, Yair Carmon, Arun Jambulapati, Yujia Jin, Aaron Sidford

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical Our paper demonstrates that our proposed optimum estimator is a useful proof device: it allows us to easily prove upper bounds on the complexity of structured optimization problems, and at least in one case (minimizing the maximum loss) improve over previously known bounds. However, our work does not investigate the practicality of our optimum estimator, as implementation and experiments are outside its scope.
Researcher Affiliation Academia Stanford University, {asi,jmblpati,yujiajin,sidford}@stanford.edu Tel Aviv University, ycarmon@tauex.tau.ac.il
Pseudocode Yes Algorithm 1: OPTEST(...) Algorithm 2: MORGRADEST(...) Algorithm 3: Stochastic accelerated gradient descent on the Moreau envelope Algorithm 4: Stochastic accelerated proximal point method Algorithm 5: Stochastic composite accelerated gradient descent Algorithm 6: Differentially-private stochastic convex optimization via optimum estimation
Open Source Code No Our paper demonstrates that our proposed optimum estimator is a useful proof device: it allows us to easily prove upper bounds on the complexity of structured optimization problems, and at least in one case (minimizing the maximum loss) improve over previously known bounds. However, our work does not investigate the practicality of our optimum estimator, as implementation and experiments are outside its scope.
Open Datasets No The paper focuses on theoretical contributions and does not conduct empirical studies, thus it does not refer to publicly available datasets or provide access information for training data.
Dataset Splits No The paper focuses on theoretical contributions and does not conduct empirical studies, thus it does not describe train/validation/test dataset splits.
Hardware Specification No Our paper demonstrates that our proposed optimum estimator is a useful proof device: it allows us to easily prove upper bounds on the complexity of structured optimization problems, and at least in one case (minimizing the maximum loss) improve over previously known bounds. However, our work does not investigate the practicality of our optimum estimator, as implementation and experiments are outside its scope.
Software Dependencies No The paper does not detail specific software dependencies with version numbers, as it is a theoretical work and does not include an experimental section requiring such specifications.
Experiment Setup No Our paper demonstrates that our proposed optimum estimator is a useful proof device: it allows us to easily prove upper bounds on the complexity of structured optimization problems, and at least in one case (minimizing the maximum loss) improve over previously known bounds. However, our work does not investigate the practicality of our optimum estimator, as implementation and experiments are outside its scope.