NSF supports More Modular Deep Learning
October 01, 2022
The National Science Foundation is supporting the lab’s project on more modular deep learning. The work builds on our finding that a trained neural network can be decomposed into modules, and asks how to make those modules more general, more reliable, and easier to reuse and replace.
Modularity offers a path to building and updating models at lower cost and energy, and to repairing them by replacing the part responsible for a problem rather than retraining the whole network.
This work is part of Modular and Dependable AI; see our modular deep learning overview. Details are on the NSF award page.