What is it about?
Deep neural networks are becoming more prevalent in aerospace engineering, in real-time applications and in numerical simulations. These networks contain many parameters that must be optimized through a training process on existing data, a challenging task given the size of the networks. Parameter-adaptive techniques can improve training robustness and efficiency, similarly to mesh-adaptation for partial differential equations. This paper presents such a technique for an ordinary-differential equation network architecture.
Featured Image
Photo by JJ Ying on Unsplash
Why is it important?
We demonstrate efficiency improvements in adaptive versus direct training of a large network. Since deep neural networks are becoming more widely used, reductions in training time can enable larger networks, less wasted CPU power, and improved network performance.
Perspectives
While machine learning has influenced recent research in computational fluid dynamics (CFD), this work tries to take ideas in the other direction: from CFD to machine learning.
Krzysztof Fidkowski
University of Michigan
Read the Original
This page is a summary of: Adjoint-Based Adaptive Training of Deep Neural Networks, July 2021, American Institute of Aeronautics and Astronautics (AIAA),
DOI: 10.2514/6.2021-2904.
You can read the full text:
Contributors
The following have contributed to this page
