What is it about?
Neural networks often require large amounts of training data. This can be prohibitive for certain applications in science and engineering. We develop methods which train neural networks using data from multiple models of different qualities and costs, yielding better results than if you had used a single high quality model. Since we are also working with high dimensional data we use various techniques to reduce their dimensionality and handle it within the network.
Featured Image
Photo by Sergey Svechnikov on Unsplash
Why is it important?
Neural networks can now be used for even more applications, even when limited by data or dimensionality. If multi-fidelity data is available, the methods presented were shown to be superior to alternate approaches for the numerical benchmarks tested.
Perspectives
This was a great collaboration between many institutions and authors. Our colleagues took great effort to generate the data for us and develop the best approaches for this task.
Vignesh Sella
University of Texas at Austin
Read the Original
This page is a summary of: Improving Neural Network Efficiency With Multifidelity and Dimensionality Reduction Techniques, January 2025, American Institute of Aeronautics and Astronautics (AIAA),
DOI: 10.2514/6.2025-2807.
You can read the full text:
Resources
Contributors
The following have contributed to this page
