Closing the convergence gap of SGD without replacement

Authors: Shashank Rajput, Anant Gupta, Dimitris Papailiopoulos

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To verify our lower bound of Theorem 2, we ran SGDo on the function described in Eq. (5) with L = 4. The step size regimes that were considered were α = 1 T , 2 log T T , 4 log T T , 8 log T n. The plot for α = 4 log T T is shown in Figure 2.
Researcher Affiliation Academia 1University of Wisconsin Madison. Correspondence to: Shashank Rajput <rajput3@wisc.edu>.
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes The code for these experiments is available at https://github.com/shashankrajput/SGDo.
Open Datasets No The paper uses a custom function described in Eq. (5) for numerical verification and does not mention or provide access to a publicly available or open dataset. 'To verify our lower bound of Theorem 2, we ran SGDo on the function described in Eq. (5) with L = 4.'
Dataset Splits No The paper describes experiments on a constructed function by varying parameters (K, n) but does not specify training, validation, or test dataset splits in the conventional sense.
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU models, or memory specifications used for running experiments.
Software Dependencies No The paper does not list any specific software dependencies with version numbers (e.g., programming languages, libraries, or solvers).
Experiment Setup Yes The step size regimes that were considered were α = 1 T , 2 log T T , 4 log T T , 8 log T n. ... For each value of K, say K = 50, we set α = T = 4 log(n K) n K = 4 log(500 50) 500 50 and ran SGDo with this constant step size α on the sum of n = 500 functions for K = 50 epochs, and the final error was recorded. ... The optimization was initialized at the origin, that is x1 0 = 0.