Sign In

Communications of the ACM

ACM TechNews

Deep Learning Techniques Lead to Materials Imaging Breakthrough

View as: Print Mobile App Share:
An image of scanning transmission electron microscope data.

The team’s techniques dramatically increased the number of images that can be processed at once while training deep neural networks.

Credit: Junqi Yin/Oak Ridge National Laboratory

Researchers at the U.S. Department of Energy's Oak Ridge National Laboratory (ORNL), software company Nvidia, and transport firm Uber Technologies have created two new deep learning methods.

The team used Uber's Horovod deep neural network (DNN) training platform, eliminating repetitive steps to boost training speed with the new methods.

The first technique uses a response cache that stores metadata from each operator request in Horovod to reduce coordinator-worker communication; the second improves power efficiency by grouping mathematical operations of multiple DNN models by exploiting similarities in each model's calculations.

Combining the methods almost doubled Horovod's scaling efficiency, which the team quantified by running the resulting STEMDL code on the IBM AC922 Summit supercomputer.

Researchers used the STEMNDL network to solve a persistent materials-imaging inverse problem involving precise analysis of scanning transmission electron microscope data.

From Oak Ridge National Laboratory
View Full Article


Abstracts Copyright © 2022 SmithBucklin, Washington, DC, USA


No entries found

Sign In for Full Access
» Forgot Password? » Create an ACM Web Account