Disentangled Multi-Fidelity Deep Bayesian Active Learning

Authors: Dongxia Wu, Ruijia Niu, Matteo Chinazzi, Yian Ma, Rose Yu

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

Reproducibility Variable Result LLM Response
Research Type Experimental On benchmark tasks of learning deep surrogates of partial differential equations including heat equation, Poisson s equation and fluid simulations, our approach significantly outperforms state-of-the-art in prediction accuracy and sample efficiency. (Abstract) [...] 5. Experiments
Researcher Affiliation Academia 1Department of Computer Science and Engineering, University of California San Diego, La Jolla, USA 2Halıcıo glu Data Science Institute, University of California San Diego, La Jolla, USA 3The Roux Institute, Northeastern University, Portland, USA 4Network Science Institute, Northeastern University, Boston, USA.
Pseudocode Yes Algorithm 1 Batch MF-LIG
Open Source Code Yes Our code is available at https://github.com/Rose-STLLab/Multi-Fidelity-Deep-Active-Learning.
Open Datasets No The paper describes how the ground-truth data is 'generated from the numerical solver' or 'simulated' based on cited equations (Olsen-Kettle, 2011; Holl et al., 2020), but does not provide direct access (link, repository, or formal citation) to these generated datasets as pre-existing public datasets.
Dataset Splits No For both passive and active learning, we randomly generate 512 data points as the test set for 4 benchmark tasks and 256 data points as the test set for fluid simulation. (Section 5.2) The paper does not specify a validation set split.
Hardware Specification No The paper does not provide specific details regarding the hardware used for running the experiments (e.g., GPU/CPU models, memory, or cloud instance types).
Software Dependencies No The paper does not provide specific software dependency details with version numbers (e.g., Python 3.x, PyTorch 1.x).
Experiment Setup Yes For active learning, we use the same 8 uniformly sampled data points across all fidelity levels as the reference data for initial training. We run 25 iterations and at each iteration, the active learning framework queries the simulator for the input with the highest acquisition function score until it reaches the budget limit of 20 per iteration.