Streaming Bayes GFlowNets

Authors: Tiago Silva, Daniel Augusto de Souza, Diego Mesquita

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our case studies in linear preference learning and phylogenetic inference showcase the effectiveness of SB-GFlow Nets in sampling from an unnormalized posterior in a streaming setting. As expected, we also observe that SB-GFlow Nets is significantly faster than repeatedly training a GFlow Net from scratch to sample from the full posterior. 5 Experiments We show that SB-GFlow Nets can accurately learn the posterior distribution conditioned on the streaming data for one toy and two practically relevant applications.
Researcher Affiliation Academia Tiago da Silva Getulio Vargas Foundation tiago.henrique@fgv.br Daniel Augusto de Souza University College London daniel.souza.21@ucl.ac.uk Diego Mesquita Getulio Vargas Foundation diego.mesquita@fgv.br
Pseudocode Yes Algorithm 1 Training a SB-GFlow Net by minimizing LSB
Open Source Code Yes Answer: [Yes] Justification: Provided in a zip file.
Open Datasets No We generate the data by simulating JC69 s model for a collection of N = 7 species and a substitution rate of λ = 5 10 1 (see [59], Chapter 1). At each iteration, we sample a new DNA subsequence of size 102 for each species and update SB-GFlow Net according to Algorithm 1. We assume that x [[0, 4]]d and d = 24 and that the data was simulated from the observational model. The paper uses simulated/generated data for its experiments and does not provide links or citations to pre-existing, publicly available datasets.
Dataset Splits No The paper describes a streaming data setting where data is continuously collected and processed. It does not explicitly mention traditional train/validation/test dataset splits with percentages or counts for its experiments.
Hardware Specification Yes Experiments were run in a cluster equipped with A100 and V100 GPUs, using a single GPU per run.
Software Dependencies No To approximately solve the optimization problem outlined in Algorithm 1, we employed the Adam optimizer [27] with a learning rate of 10 3 for the p F s parameters and 10 1 for log Zt, following recommendations from [37]. It mentions the optimizer but no specific software library versions (e.g., PyTorch, TensorFlow version).
Experiment Setup Yes We provide below details for reproducing our experiments for each considered generative task. To approximately solve the optimization problem outlined in Algorithm 1, we employed the Adam optimizer [27] with a learning rate of 10 3 for the p F s parameters and 10 1 for log Zt, following recommendations from [37]. Also, we linearly decreased the learning rate during training.