Improved SVRG for Non-Strongly-Convex or Sum-of-Non-Convex Objectives
Authors: Zeyuan Allen-Zhu, Yang Yuan
ICML 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We confirm our theoretical findings using four real-life datasets: (1) the Adult dataset (32, 561 examples and 123 features), (2) the Covtype dataset (581, 012 examples and 54 features), (3) the Ijcnn1 dataset (49990 examples and 22 features), and (4) the 2nd class of the MNIST dataset (60, 000 examples and 780 features) (Fan & Lin). |
| Researcher Affiliation | Academia | Zeyuan Allen-Zhu ZEYUAN@CSAIL.MIT.EDU Princeton University Yang Yuan YANGYUAN@CS.CORNELL.EDU Cornell University |
| Pseudocode | Yes | Algorithm 1 SVRG++(xφ, m0, S, η) Algorithm 2 SVRG(xφ, m, S, η) (Johnson & Zhang, 2013) |
| Open Source Code | No | The paper does not provide any links to open-source code repositories or explicitly state that the source code for the described methodology is available. |
| Open Datasets | Yes | We confirm our theoretical findings using four real-life datasets: (1) the Adult dataset (32, 561 examples and 123 features), (2) the Covtype dataset (581, 012 examples and 54 features), (3) the Ijcnn1 dataset (49990 examples and 22 features), and (4) the 2nd class of the MNIST dataset (60, 000 examples and 780 features) (Fan & Lin). |
| Dataset Splits | No | The paper mentions 'training loss' and describes tuning parameters, but it does not specify explicit training/validation/test dataset splits (e.g., percentages, sample counts, or cross-validation details). |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for conducting the experiments (e.g., CPU/GPU models, memory specifications, or cloud computing instances). |
| Software Dependencies | No | The paper lists algorithms implemented (SVRG++, SAGA, SDCA) but does not specify any software frameworks, libraries, or their version numbers used in the implementation or for experimental setup. |
| Experiment Setup | Yes | For each algorithm above except SDCA, we tune the step length carefully from the set {a 10 k : a {1, 2, . . . , 9}, k Z} for each plot. SVRG++ with initial epoch length m0 = n/4. |