2018 R&D 100 Award for Multinode Evolutionary Neural Networks for Deep Learning (MENNDL)

The MENNDL deep-learning technology was designed to mimic human thought processes. When analyzing a dataset, the software starts with a poorly performing network, then alters it through a series of feedback loops to optimize its performance within hours.

MENNDL’s evolving networks are applicable in many fields from computer vision to speech recognition and can be trained to analyze specific datasets. The system presents a novel application of deep learning algorithms to electron microscopy, as it is able to extract structural data from raw atomic microscopy information.

MENNDL is also a potentially more efficient approach to data analysis, as its custom-generated neural networks can match or exceed the performance of handcrafted networks. The system has run on the Summit supercomputer, but future developments will allow MENNDL networks to run on smaller computers.

The ORNL development team was led by Robert Patton and included Thomas Karnowski, Seung-Hwan Lim, Thomas Potok, Derek Rose and Steven Young.

MENNDL was funded through ORNL’s Laboratory Directed Research and Development Program.