Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
Multi-Fidelity Active Learning with GFlowNets
Authors: Alex Hernández-García, Nikita Saxena, Moksh Jain, Cheng-Hao Liu, Yoshua Bengio
TMLR 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we present empirical evaluations of multi-fidelity active learning with GFlow Nets. Through our experiments, we aim to answer the following questions: Can our multi-fidelity active learning approach find high-scoring, diverse samples at lower cost than with a single-fidelity oracle? |
| Researcher Affiliation | Academia | Alex Hernandez-Garcia EMAIL Mila, Université de Montréal Nikita Saxena* EMAIL Birla Institute of Technology and Science, Pilani Moksh Jain EMAIL Mila, Université de Montréal Cheng-Hao Liu EMAIL Mila, Mc Gill University Yoshua Bengio EMAIL Mila, Université de Montréal, CIFAR Fellow, IVADO |
| Pseudocode | Yes | Algorithm 1: MF-GFN: Multi-fidelity active learning with GFlow Nets. A graphical summary of this algorithm is shown in Fig. 1. |
| Open Source Code | Yes | Finally, the implementation of MF-GFN and the code to reproduce the experiments is publicly available in a Git Hub repository: https://github.com/nikita-0209/mf-al-gfn |
| Open Datasets | Yes | We use data from DBAASP (Pirtskhalava et al., 2021), containing antimicrobial activity labels, which is originally split into three sets: D1 for training the oracle, D2 as the initial data set in the active learning loop and D3 as the test set (Jain et al., 2022). |
| Dataset Splits | Yes | The initial data set is split into train and validation in the ratio of 9:1 for all tasks. |
| Hardware Specification | No | The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | Yes | We implement the oracles using RDKit 2023.03 (rdk, 2023) and the semi-empirical quantum chemistry package x TB. |
| Experiment Setup | Yes | Table 2: Hyperparameters concerning the active learning setting and the policy reward function. Table 4: Deep kernel hyperparameters for the DNA and antimicrobial tasks. Neural Network Architecture For all experiments, the same base architecture was used, featuring transformer encoder layers with position masking for padding tokens. Standard pre-activation residual blocks were included, comprising two convolutional layers, layer normalisation, and swish activations. Optimiser Hyperparameters The running estimates of the first two moments in the Adam optimiser (Kingma & Ba, 2015) were disabled by setting β1 = 0.0 and β2 = 0.01. |