GFlowNets and variational inference
Authors: Nikolay Malkin, Salem Lahlou, Tristan Deleu, Xu Ji, Edward J Hu, Katie E Everett, Dinghuai Zhang, Yoshua Bengio
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate that, in certain cases, VI algorithms are equivalent to special cases of GFlow Nets in the sense of equality of expected gradients of their learning objectives. We then point out the differences between the two families and show how these differences emerge experimentally. |
| Researcher Affiliation | Collaboration | Nikolay Malkin , Salem Lahlou , Tristan Deleu Mila, Universit e de Montr eal Xu Ji, Edward Hu Mila, Universit e de Montr eal Katie Everett Google Research Dinghuai Zhang Mila, Universit e de Montr eal Yoshua Bengio Mila, Universit e de Montr eal, CIFAR |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. Appendix A contains a figure illustrating a concept, not an algorithm. |
| Open Source Code | Yes | Code: https://github.com/GFNOrg/GFN_vs_HVI. |
| Open Datasets | No | The paper mentions using a 'synthetic hypergrid environment introduced by Bengio et al. (2021a)' and that 'for each experiment, we sampled a dataset D of 100 observations'. While these refer to data, no concrete access information (link, DOI, specific citation for the dataset itself) is provided for public access to the datasets used in the experiments. |
| Dataset Splits | No | The paper mentions 'a held-out set' for molecule experiments but does not specify exact split percentages, absolute sample counts, or reference predefined splits with citations for training, validation, or test sets across all experiments. |
| Hardware Specification | No | The paper states: 'This research was enabled in part by computational resources provided by the Digital Research Alliance of Canada.' This is a general statement and does not provide specific hardware details (e.g., CPU/GPU models, memory). |
| Software Dependencies | No | The paper mentions using 'Adam optimizer' and 'SGD' but does not provide specific version numbers for these software components or any other libraries/packages used. |
| Experiment Setup | Yes | The forward and backward policies are parametrized as neural networks with 2 hidden layers of 256 units each. We use a batch size of 64 for all learning objectives... The Adam optimizer... All models were trained with the Adam optimizer and batch size 4 for a maximum of 50000 batches. We used the Adam optimizer, with the best learning rate found among {10 6, 3 10 6, 10 5, 3 10 5, 10 4}. For the TB objective, we learned log š¯‘¨using SGD with a learning rate of 0.1 and momentum 0.8. |