Generative Flow Networks for Discrete Probabilistic Modeling

Authors: Dinghuai Zhang, Nikolay Malkin, Zhen Liu, Alexandra Volokhova, Aaron Courville, Yoshua Bengio

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate EB-GFN s effectiveness on various probabilistic modeling tasks. Code is publicly available at github.com/zdh Narsil/EB GFN. and We test the algorithm on a variety of synthetic and real tasks, achieving competitive results ( 4).
Researcher Affiliation Academia 1Mila Quebec AI Institute and Universit e de Montr eal, Montreal, Quebec, Canada.
Pseudocode Yes Algorithm 1 EB-GFN joint training framework and Algorithm 2 GFlow Net-guided energy function update
Open Source Code Yes Code is publicly available at github.com/zdh Narsil/EB GFN.
Open Datasets Yes Omniglot, Silhouettes, Static MNIST, Dynamic MNIST are mentioned in Table 3. These are standard, well-known public datasets used in machine learning.
Dataset Splits Yes The validation and evaluation protocol is also kept aligned with Grathwohl et al. (2021b), where the checkpoint with the best negative log-likelihood on the validation set is reported. and The validation of EBM likelihood is achieved with 300000 step annealed importance sampling (AIS).
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory specifications, or cloud resources) used for running its experiments.
Software Dependencies No The paper mentions software components like 'Adam optimizer' and 'ELU activation' but does not provide specific version numbers for any software dependencies, libraries, or programming languages.
Experiment Setup Yes For simplicity, we set K = D for the negative sampling step and use a training policy with α = 1 (no backward paths from training examples). Details can be found in C.1. and We use a 4 layer MLP with 256 hidden dimension and ELU activation (Clevert et al., 2016) as the energy function. The training of this energy function lasts 105 steps. An Adam optimizer with 1 10 3 learning rate is used to update the EBM. The batch size is 128. and For GFlow Net training, we use a mixed training policy with α = 0.5 and a schedule with linearly increasing K for the back-and-forth proposal. See C.2 for details.