Multi Fidelity Active Learning

Multi-fidelity active learning aims to efficiently train machine learning models using data from multiple sources with varying costs and accuracies (fidelities). Research focuses on developing algorithms that strategically select which data points, and at what fidelity level, to acquire, maximizing model accuracy while minimizing the overall cost of data acquisition. Prominent approaches leverage Bayesian methods, deep learning architectures (including neural operators and GFlowNets), and novel acquisition functions designed to balance exploration and exploitation across fidelities. This approach is particularly impactful for computationally expensive simulations in scientific and engineering domains, accelerating model development and reducing resource consumption.