Leveraging Recursive Gumbel-Max Trick for Approximate Inference in Combinatorial Spaces
Authors: Kirill Struminsky, Artyom Gadetsky, Denis Rakitin, Danil Karpushkin, Dmitry P. Vetrov
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In our experiments, we consider various structured latent variable models and achieve results competitive with relaxation-based counterparts. ... 5 Applications ... Table 1: Results of k-subset selection on Aroma aspect data. ... Table 2: Graph Layout experiment results for T=10 iterations. ... Table 3: Unsupervised Parsing on List Ops. |
| Researcher Affiliation | Collaboration | Kirill Struminsky HSE University Moscow, Russia k.struminsky@gmail.com Artyom Gadetsky HSE University Moscow, Russia artygadetsky@yandex.ru Denis Rakitin HSE University, Skoltech Moscow, Russia rakitindenis32@gmail.com Danil Karpushkin AIRI, Sber AI Lab, MIPT Moscow, Russia kardanil@mail.ru Dmitry Vetrov HSE University, AIRI Moscow, Russia vetrovd@yandex.ru |
| Pseudocode | Yes | Figure 1: The recursive algorithm for arg top k and the general algorithm with the stochastic invariant put side-by-side. ... Algorithm 1 Ftop-k(E, K, k) ... Algorithm 2 Fstruct(E, K, R) ... Algorithm 3 Flog-prob(t, λ, K, R) |
| Open Source Code | Yes | The implementation and how-to-use examples are publicly available8. https://github.com/RakitinDen/pytorch-recursive-gumbel-max-trick |
| Open Datasets | Yes | We evaluated our method on the experimental setup from L2X [2] on the Beer Advocate [30] dataset. ... Following details about data and models outlined by [34], we use a simplified version of the List Ops [32] dataset. |
| Dataset Splits | No | Table 2: Graph Layout experiment results for T=10 iterations. Metrics are obtained by choosing models with best validation ELBO and averaging results across different random seeds on the test set. No explicit details about validation dataset split percentages or sizes were found in the provided text. |
| Hardware Specification | No | The research was supported by the Russian Science Foundation grant no. 19-71-30020 and through the computational resources of HPC facilities at HSE University[23]. No specific hardware details (e.g., GPU/CPU models, memory) were explicitly mentioned for running experiments. |
| Software Dependencies | No | The implementation and how-to-use examples are publicly available8. https://github.com/RakitinDen/pytorch-recursive-gumbel-max-trick. No specific software dependencies with version numbers were mentioned. |
| Experiment Setup | Yes | Following setup of [34] we use k = {5, 10, 15} and two CNN architectures: Simple a one-layer CNN and Complex a three-layer CNN and train our models for each aspect using MSE loss. ... Detailed experimental setup and description of models can be found in Appendix D.1. |