A theory of continuous generative flow networks

Authors: Salem Lahlou, Tristan Deleu, Pablo Lemos, Dinghuai Zhang, Alexandra Volokhova, Alex Hernández-Garcı́a, Lena Nehale Ezzine, Yoshua Bengio, Nikolay Malkin

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4. Experiments
Researcher Affiliation Academia 1Mila 2Universit e de Montr eal 3Ciela Institute 4CIFAR.
Pseudocode No The paper does not contain any pseudocode or algorithm blocks.
Open Source Code Yes Code for these experiments can be found at https://github. com/saleml/continuous-gfn.
Open Datasets Yes We evaluate GFlow Nets and baselines on two synthetic densities: a 2-dimensional mixture of 9 Gaussians and the 10-dimensional funnel from MCMC literature (Hoffman & Gelman, 2011)." and "We train a GFlow Net as described above on the Image Net-32 dataset (treated as a set of terminating states)
Dataset Splits No The paper discusses training and testing but does not explicitly provide details about validation dataset splits (percentages, counts, or methodology).
Hardware Specification Yes This is much shorter than other state-of-the-art work (such as Lipman et al. (2022)) and takes less than 3 days on a single V100 GPU.
Software Dependencies No The paper does not provide specific software dependencies with version numbers.
Experiment Setup Yes We learned the concentration parameters of the Beta distributions, which were restrained to the interval [0.1, 5.1], using a three-layered neural network with 128 units per layer, and leaky Re LU activation for 𝑠 𝑠0...We trained the models for 20,000 iterations. For both the DB and TB losses, we used a learning rate of 10 3 for the parameters of 𝑝𝐹, 𝑝𝐵, 𝑝𝐹(𝑠0, ) (and log 𝑍for TB, 𝑢for DB). The learning rate was annealed using a discount factor of 0.5 every 2500 iterations.