Embarrassingly Parallel GFlowNets
Authors: Tiago Silva, Luiz Max Carvalho, Amauri H Souza, Samuel Kaski, Diego Mesquita
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments validate EP-GFlow Nets in different contexts, including multi-objective multiset, parallel Bayesian phylogenetic inference, and federated Bayesian structure learning. In summary, our contributions are: 1. We propose EP-GFlow Net, the first algorithm for embarrassingly parallel sampling in discrete state spaces using GFlow Nets. We provide theoretical guarantees of correctness, and also analyze EP-GFlow Net s robustness to errors in the estimation of local GFlow Nets; 2. We present the contrastive balance (CB) condition, show it is a sufficient and necessary condition for sampling proportionally to a reward, and analyze its connection to variational inference (VI); 3. We substantiate our methodological contributions with experiments on five different tasks. Our empirical results i) showcase the accuracy of EP-GFlow Nets; ii) show that, in some cases, using the CB as training criterion leads to faster convergence compared to popular loss functions; iii) illustrate EP-GFlow Nets potential in two notable applications: Bayesian phylogenetic inference, and Bayesian network structure learning. |
| Researcher Affiliation | Collaboration | 1Getulio Vargas Foundation 2Aalto University 3Federal Institute of Cear a 4University of Manchester. Correspondence to: Diego Mesquita <diego.mesquita@fgv.br>. |
| Pseudocode | Yes | Algorithm 1 Training of EP-GFlow Nets |
| Open Source Code | Yes | The computer code for reproducing our experiments will be publicly released at github.com/ML-FGV/ep-gflownets. |
| Open Datasets | No | The paper uses standard benchmark datasets for tasks like 'multiset generation' and 'design of sequences', and describes custom data generation for 'grid world', 'Bayesian phylogenetic inference', and 'federated Bayesian network structure learning'. However, for the custom datasets, it does not provide concrete access information (links, DOIs, formal citations for the specific generated datasets) to allow direct public access. For example, 'We designed the GFlow Net to generate multisets of size 8 by iteratively selecting elements from a set U of size 10.' without providing a way to access that specific generated data or the generation script. |
| Dataset Splits | No | The paper mentions training and inference, but does not specify explicit train/validation/test dataset splits, percentages, or absolute counts for any of its tasks. For example, it says 'For inference, we simulated 10^6 environments' for the grid world, but doesn't detail how the data was split for training. |
| Hardware Specification | Yes | The reported times were measured in a high-performance Linux machine equipped with an AMD EPYC 7V13 64-core processor, 216 GB DDR4 RAM and a NVIDIATM A100 80 GB PCIe 4.0 GPU. |
| Software Dependencies | No | The paper mentions using 'Py Torch' and 'Py Torch Geometric' but does not specify their version numbers. It also refers to 'Adam W optimizer (Loshchilov & Hutter, 2019)' and 'Leaky ReLU activation function (Maas et al., 2013)', but these are algorithms/functions, not specific software dependencies with version numbers. |
| Experiment Setup | Yes | For the stochastic optimization, we minimized the contrastive balance objective using the Adam W optimizer (Loshchilov & Hutter, 2019) for both local and global GFlow Nets. We trained the models for 5000 epochs (20000 for the grid world) with a learning rate equal to 3 x 10^-3 with a batch size dependent upon the environment. |