Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
On Convergence of Distributed Approximate Newton Methods: Globalization, Sharper Bounds and Beyond
Authors: Xiao-Tong Yuan, Ping Li
JMLR 2020 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Numerical evidence is provided to confirm the theoretical and practical advantages of our methods. Numerical evaluation results are presented and discussed in Section 4. Finally, we conclude this article in Section 5. All the technical proofs of theoretical results are deferred to the appendix section. |
| Researcher Affiliation | Industry | Xiao-Tong Yuan EMAIL Cognitive Computing Lab Baidu Research Beijing 100193, China Ping Li EMAIL Cognitive Computing Lab Baidu Research Bellevue, WA 98004, USA |
| Pseudocode | Yes | Algorithm 1: DANE with backtracking Line Search: DANE-LS(γ, ρ, ν) Algorithm 2: DANE with Heavy-Ball acceleration: DANE-HB(γ, β) Algorithm 3: Distributed Doubly Approximate Newton: D2ANE(γ, β, ℓ) |
| Open Source Code | No | The paper does not contain an explicit statement about the release of their own source code, nor does it provide a link to a code repository for the described methodology. It mentions using 'SGDLibrary (Kasai, 2017)' as a third-party solver. |
| Open Datasets | Yes | Next, we evaluate the convergence performance of the considered algorithms on two real data sets gisette (Guyon et al., 2005) (p = 5000, N = 6000) and rcv1.binary (Lewis et al., 2004) (p = 47236, N = 20242). |
| Dataset Splits | No | We replicate each experiment 10 times over random split of data and report the results in mean-value along with error bar. |
| Hardware Specification | Yes | We simulate the distributed environment on a single server powered by dual Intel(R) Xeon(R) E5-2630V4@2.2GHz CPU with multiple logic processors simulating multiple machines. |
| Software Dependencies | Yes | All the considered methods are implemented in Matlab R2018b on Microsoft Windows 10. The local subproblems on the master machine are solved by an SVRG solver from SGDLibrary (Kasai, 2017) |
| Experiment Setup | Yes | We initialize w(0) = 0 throughout our numerical study. For our simulation study, we test with feature dimensions p {200, 500}. We fix N = 10p, µ = 1/ N, and study the impact of varying number of machines m and regularization γ = O(1/ n) on the needed rounds of communication to reach sub-optimality ϵ = 10 6. For each data set, we fix the regularization parameter µ = 10 5 and test with m {4, 16, 32}. |