Revisiting Zeroth-Order Optimization: Minimum-Variance Two-Point Estimators and Directionally Aligned Perturbations

Authors: Shaocong Ma, Heng Huang

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through empirical evaluations on both synthetic problems and practical tasks, we demonstrate that DAPs outperform traditional methods under specific conditions.
Researcher Affiliation Academia Shaocong Ma and Heng Huang Department of Computer Science University of Maryland College Park, MD 20742, USA EMAIL
Pseudocode Yes Algorithm 1: The algorithm for sampling from a hyper-plane Algorithm 2: A practical implementation of gradient estimator using DAPs
Open Source Code Yes All source codes, including visualization scripts, are provided with our submission.
Open Datasets Yes We conducted experiments using the OPT-1.3b model (Zhang et al., 2022) for sentiment classification on the Stanford Sentiment Treebank (SST-2) dataset (Socher et al., 2013).
Dataset Splits Yes We conducted experiments using the OPT-1.3b model (Zhang et al., 2022) for sentiment classification on the Stanford Sentiment Treebank (SST-2) dataset (Socher et al., 2013).
Hardware Specification Yes We conducted our experiments on a cluster running RHEL8, equipped with Dual AMD EPYC 9124 processors and eight NVIDIA RTX 6000 Ada Generation graphics cards.
Software Dependencies No Our code was tested using Python version 3.10.10. Additional dependencies are specified in the supplementary requirements.txt file.
Experiment Setup Yes Learning rate η: 10 4; Perturbation size µ: 10 5; Zeroth-order gradient estimation batch size b: 2; Stochastic gradient updates batch size: 16.