Presentation
SIGN IN TO VIEW THIS PRESENTATION Sign In
Toward Rapid Autonomous Electron Microscopy with Active Meta-Learning
DescriptionIn this work, we developed a method to accelerate computational steering of microscopy experiments by active meta-learning. Before this work, a tailored AI model was trained specifically for every experiment by active learning to reconstruct spectrum and uncover regions of interest by sampling just a few locations of the image. Training individual models for each experiment may result in scalability challenges when dealing with high resolutions data, and complex structure-property relationships often demand deeper AI models. A Reptile algorithm, a first-order, model-agnostic meta-learning approach is used to train on images from prior experiments at different conditions such that the trained model can adapt to new unseen conditions in considerably less time. We observe up to ~30-40% reduction in the number of training epochs for active learning exploration. The benefit for structure-property investigation for spectral reconstruction of STEM EELS nanoparticle plasmonic images is demonstrated across multiple experiments.
Event Type
Workshop
TimeMonday, 13 November 20234:10pm - 4:30pm MST
Location501-502
Artificial Intelligence/Machine Learning
W