Physarum Powered Differentiable Linear Programming Layers and Applications

Authors: Zihang Meng, Sathya N. Ravi, Vikas Singh8939-8949

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

Reproducibility Variable Result LLM Response
Research Type Experimental We describe our development and show the use of our solver in a video segmentation task and meta-learning for few-shot learning. Our solver performs comparably with a customized projected gradient descent method on the first task and outperforms the differentiable CVXPY-SCS solver on the second task. Experiments show that our solver converges quickly without the need for a feasible initial point.
Researcher Affiliation Academia Zihang Meng, 1 Sathya N. Ravi, 2 Vikas Singh 1 1 University of Wisconsin-Madison 2 University of Illinois at Chicago zihangm@cs.wisc.edu, sathya@uic.edu, vsingh@biostat.wisc.edu
Pseudocode Yes Algorithm 1: γ Aux PD Layer
Open Source Code Yes Our code is available at https://github.com/zihangm/Physarum-Differentiable-LP-Layer and integration with CVXPY is ongoing, which will complement functionality offered by tools like Opt Net and CVXPY-SCS.
Open Datasets Yes Experiments on Youtube-VOS. Parameter settings. Datasets. We follow the code from (Lee et al. 2019) to conduct the experiments on CIFAR-FS and FC100.
Dataset Splits Yes We use Ngrad = 40, Nproj = 5, lr = 0.1 as in their paper to reproduce their results. For γ Aux PD layer, the choice is simple: step size h = 1 and K = 10 iterations work well for both two experiments and the other tests we performed. The results on CIFAR-FS and FC100 are shown in Table 3. Using the ℓ1 normalized SVM, our solver achieves better performance than CVXPY-SCS (Agrawal et al. 2019) and Optnet (with a small quadratic term as regularization) on both datasets and both the 1-shot and 5-shot setting.
Hardware Specification No The paper mentions "on GPU" in Table 4, but does not specify any particular GPU models, CPU models, or other detailed hardware specifications for running the experiments.
Software Dependencies No The paper mentions using the "CVXPY package" and refers to "Opt Net" but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes Parameter settings. The projection gradient descent solver in (Zeng et al. 2019) has three parameters to tune: number of gradient steps, number of projections, and learning rate. We use Ngrad = 40, Nproj = 5, lr = 0.1 as in their paper to reproduce their results. For γ Aux PD layer, the choice is simple: step size h = 1 and K = 10 iterations work well for both two experiments and the other tests we performed. The implementation of Optnet (Amos and Kolter 2017) does not directly support solving LPs since it requires a positive definite quadratic term. Still, to test its ability of solving LPs, we add a diagonal matrix with a small value (0.1, since diagonal value smaller than 0.1 leads to numerical errors in our experiment) as the quadratic term (can be thought of as a regularization term).