Multi-Objective GFlowNets
Authors: Moksh Jain, Sharath Chandra Raparthy, Alex Hernández-Garcı́a, Jarrid Rector-Brooks, Yoshua Bengio, Santiago Miret, Emmanuel Bengio
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we present our empirical findings across a wide range of tasks ranging from sequence design to molecule generation. Through our experiments, we aim to answer the following questions: Q1 Can MOGFNs model the preference-conditional reward distribution? Q2 Can MOGFNs sample Pareto-optimal candidates? Q3 Are candidates sampled by MOGFNs diverse? Q4 Do MOGFNs scale to high-dimensional problems relevant in practice? We obtain positive experimental evidence for Q1-Q4. |
| Researcher Affiliation | Collaboration | Moksh Jain 1 2 Sharath Chandra Raparthy 1 2 * Alex Hernandez-Garcia 1 2 Jarrid Rector-Brooks 1 2 Yoshua Bengio 1 2 3 Santiago Miret 4 Emmanuel Bengio 5 1Universite de Montreal 2Mila Quebec AI Institute 3CIFAR Fellow & IVADO 4Intel Labs 5Recursion. |
| Pseudocode | Yes | A. Algorithms We summarize the algorithms for MOGFN-PC and MOGFN-AL here. Algorithm 1 Training preference-conditional GFlow Nets [...] Algorithm 2 Training MOGFN-AL |
| Open Source Code | Yes | Code The code for the experiments is available at https://github.com/GFNOrg/ multi-objective-gfn. |
| Open Datasets | Yes | We first consider a small-molecule generation task based on the QM9 dataset (Ramakrishnan et al., 2014). [...] We use the dataset introduced in Stanton et al. (2022) as the initial pool of candidates D0 with |D0| = 512. |
| Dataset Splits | No | The paper mentions training models with a certain number of molecules (e.g., 'train the models with 1M molecules' for QM9) and an 'initial pool of candidates D0' for Proxy RFP, but it does not specify explicit train/validation/test dataset splits (e.g., as percentages or exact sample counts) for each of the datasets used in their experiments. While the Proxy RFP task mentions an 'initial pool' and active learning rounds, it doesn't define standard train/validation/test splits of that initial pool for model training/evaluation. |
| Hardware Specification | No | The paper mentions that 'The research was enabled in part by computational resources provided by the Digital Research Alliance of Canada (https://alliancecan.ca/en) and Mila (https: //mila.quebec),' but it does not provide specific hardware details such as exact GPU or CPU models, processor types, or memory specifications used for running the experiments. |
| Software Dependencies | No | The paper mentions several software components and libraries, such as 'RDKit library', 'NUPACK', 'Autodock Vina', 'MXMNet', 'DIAMOND', 'Fold X suite', 'Bio Python', and 'Adam optimizer'. However, it does not consistently provide specific version numbers for these software dependencies, which is required for reproducibility. |
| Experiment Setup | Yes | In Appendix D, the paper provides specific experimental setup details and hyperparameters for various tasks. For instance, in D.1 Hyper-Grid, it states: 'All models are trained with learning rate=0.01 with the Adam optimizer (Kingma & Ba, 2015) and batch size=128.' and 'We sample preferences ω from Dirichlet(α) where α = 1.5.' For N-grams, Table 6 lists 'Learning Rate (PF)', 'Learning Rate (Z)', 'Reward Exponent: β', and 'Uniform Policy Mix: δ'. Similar tables and descriptions are provided for QM9, Fragments, DNA Sequence Design, and Active Learning tasks (e.g., Table 11, 12, 13, 16). |