On the Complexity of Finite-Sum Smooth Optimization under the Polyak–Łojasiewicz Condition

Authors: Yunyan Bai, Yuxing Liu, Luo Luo

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct numerical experiments to compare DRONE with centralized gradient descent (CGD) and DGD-GT, where CGD is a distributed extension of GD in client-server networks. Please see Appendix D for details. We test the algorithms on the following three problems: Hard instance, Linear regression, Logistic regression
Researcher Affiliation Academia 1School of Data Science, Fudan University, Shanghai, China 2 Shanghai Key Laboratory for Contemporary Applied Mathematics, Shanghai, China.
Pseudocode Yes Algorithm 3 DGD-GT 1: Input: initial point x0 R1 d, iteration number T, stepsize η > 0 and communication numbers K 2: X0 = 1 x0 3: S0 = F(X0) 4: for t = 0, . . . , T 1 do 5: Xt+1 = Acc Gossip(Xt ηSt, W, K) 6: St+1 = Acc Gossip(St + F(Xt+1) F(Xt), W, K) 7: end for 8: Output: uniformly sample xout from {x T (i)}n i=1
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets Yes We evaluate the algorithms on dataset Driv Face (m = 606, d = 921, 600) (Diaz-Chito et al., 2016) for this problem. We evaluate the algorithms on dataset RCV1 (m = 20, 242, d = 47, 236) (Diaz Chito et al., 2016) for this problem.
Dataset Splits No The paper does not provide specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology).
Hardware Specification Yes All of our experiments are performed on PC with Intel(R) Core(TM) i7-8550U CPU@1.80GHz processor
Software Dependencies Yes we implement the algorithms by MPI for Python 3.9.
Experiment Setup Yes For all the above problems, we set n = 32 and use a linear graph for the network of DGD-GT and DRONE, leading to that γ = (1 cos(π/32)) / (1 + cos(π/32)) 0.0024.