An exact solver for the Weston-Watkins SVM subproblem

Authors: Yutong Wang, Clayton Scott

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform benchmark experiments on 8 datasets from LIBSVM Data: Classification (Multi-class) spanning a range of k from 3 to 1000. See Table 1. In all of our experiments, Walrus and Shark perform identically in terms of testing accuracy. We report the accuracies in Section A.6.3. Below, we will only discuss runtime. For measuring the runtime, we start the timer after the data sets have been loaded into memory and before the state variables β and w have been allocated. The primal objective is the value of (P) at the current w and the dual objective is 1 times the value of (D2) at the current β. The duality gap is the primal minus the dual objective. The objective values and duality gaps are measured after each outer iteration, during which the timer is paused.
Researcher Affiliation Academia 1Department of Electrical Engineering and Computer Science, University of Michigan. 2Department of Statistics, University of Michigan.
Pseudocode Yes Algorithm 1 Block coordinate descent on (D2), Algorithm 2 solve subproblem(v, C), Subroutine 3 get up dn seq, Subroutine 4 update vars, Subroutine 5 KKT cond.
Open Source Code Yes The solver and code for generating the figures are available4. 4See Section A.6. Our solver and code for generating figures are available at https://github.com/yutongw/walrus
Open Datasets Yes We perform benchmark experiments on 8 datasets from LIBSVM Data: Classification (Multi-class) spanning a range of k from 3 to 1000. See Table 1.
Dataset Splits No The paper mentions running experiments on datasets and evaluating duality gap, but it does not explicitly state the training, validation, and test split percentages or methodology for data partitioning.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments, such as CPU or GPU models, or memory specifications.
Software Dependencies No We implemented our linear WW-SVM subproblem solver, Walrus (Algorithm 2), along with the BCD Algorithm 1 as an extension to LIBLINEAR. We compare our implementation to Shark (Igel et al., 2008). The paper mentions specific software frameworks (LIBLINEAR, Shark) but does not provide version numbers for these or other critical dependencies (e.g., C++ compiler version, specific libraries).
Experiment Setup Yes We test our hypothesis on a larger scale by running Walrus and Shark on the datasets in Table 1 over the grid of hyperparameters C ∈ {2^6, 2^5, . . . , 2^2, 2^3}.