Black-Box Generalization: Stability of Zeroth-Order Learning

Authors: Konstantinos Nikolakakis, Farzin Haddadpour, Dionysis Kalogerias, Amin Karbasi

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical We provide the first generalization error analysis for black-box learning through derivative-free optimization. Under the assumption of a Lipschitz and smooth unknown loss, we consider the Zeroth-order Stochastic Search (Zo SS) algorithm, that updates a d-dimensional model by replacing stochastic gradient directions with stochastic differences of K + 1 perturbed loss evaluations per dataset (example) query. For both unbounded and bounded possibly nonconvex losses, we present the first generalization bounds for the Zo SS algorithm.
Researcher Affiliation Collaboration Konstantinos E. Nikolakakis Yale University konstantinos.nikolakakis@yale.edu Farzin Haddadpour Yale University farzin.haddadpour@yale.edu Dionysios S. Kalogerias Yale University dionysis.kalogerias@yale.edu Amin Karbasi Yale University & Google Research amin.karbasi@yale.edu
Pseudocode No The paper describes algorithms using mathematical equations but does not present pseudocode or clearly labeled algorithm blocks within its content.
Open Source Code No Under section '3. If you ran experiments...', all sub-sections related to code and data reproduction are marked as '[N/A]', indicating no such materials were provided.
Open Datasets No The paper is theoretical and does not describe experiments performed on a specific, publicly available dataset with concrete access information.
Dataset Splits No The paper does not report empirical studies and therefore does not provide details on training/test/validation dataset splits.
Hardware Specification No Under section '3. If you ran experiments...', the question about total amount of compute and type of resources used is marked as '[N/A]', indicating no hardware specifications were provided.
Software Dependencies No Under section '3. If you ran experiments...', the question about specifying training details (e.g., data splits, hyperparameters, how they were chosen) is marked as '[N/A]', which implies no software dependency details were provided.
Experiment Setup No Under section '3. If you ran experiments...', the question about specifying training details (e.g., data splits, hyperparameters, how they were chosen) is marked as '[N/A]', indicating no such experimental setup details were provided.