Accelerated Stochastic Optimization Methods under Quasar-convexity

Authors: Qiang Fu, Dongchu Xu, Ashia Camage Wilson

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we evaluate our methods on example (6) in the Euclidean setting using synthetic dataset... The simulation results in Figure 1 validate our methods and show the superiority of our methods in terms of the convergence speed and the overall complexity.
Researcher Affiliation Academia 1 Sun Yat-sen University, Guangzhou, China 2 Harvard University, Cambridge, MA, USA 3 MIT, Cambridge, MA, USA.
Pseudocode Yes Algorithm 1 (Ak, Bk, y0, t, ϵ)... Algorithm 2 (Ak, Bk, bk, y0, p, q, ϵ)... Algorithm 3 Bisearch(f, y, z, b, c, ϵ,[guess])
Open Source Code Yes Code is available at https://github.com/Qiang Fu09/Stochastic-quasar-convexacceleration.
Open Datasets Yes We generate the true dynamical system and data the same way as in Hardt et al. (2016) using parameters N = 5000, d = 20, T = 500. We use the following multi-classification dataset from Dua & Graff (2017), which we treat as binary classification datasets.
Dataset Splits No The paper mentions generating N training examples and using synthetic datasets or cited datasets, but it does not specify any training/validation/test dataset splits, percentages, or methodologies for partitioning data into these sets.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers, such as programming languages, libraries, or frameworks used for implementation.
Experiment Setup Yes We choose ϵ = 10^-2, the stepsize to be 5 * 10^-5, 1 * 10^-6, 1 * 10^-4 for SGD, L = 1 * 10^6, 1 * 10^8, 1 * 10^5 for QASGD and L = 3 * 10^4, 1 * 10^6, 1 * 10^4 for QASVRG in LDS1, LDS2 and LDS3. The flat line in the third column means the loss blows up to infinity with this choice of stepsize. ... We choose the value of random seed to be in {0, 12, 24, 36, 48}...