Multinode Evolutionary Neural Networks for Deep Learning (MENNDL)

Organization: Oak Ridge National Laboratory
Year: 2018

Deep learning has the potential to revolutionize scientific discovery. However, designing an optimal network topology with optimal hyperparameter values for scientific data is a challenge. Multinode Evolutionary Neural Networks for Deep Learning (MENNDL) is a scalable evolutionary algorithm, using the Oak Ridge Leadership Computing Facility’s Titan supercomputer, to design the optimal hyperparameters and topology of deep convolutional neural networks. This scalable evolutionary algorithm can automatically design a deep learning network capable of very high accuracy in classification or prediction tasks. Using this technology, researchers can now design an optimal deep learning network within a matter of hours as opposed to months and have successfully executed this algorithm on 100 percent of Titan’s 18,688 nodes for periods up to 24 consecutive hours. This technology is the first to provide researchers with the ability to optimize the interactions between parameters and their values as applied to different data sets. It also significantly reduces the expertise and knowledge required for parameter tuning. Solving the hyperparameter optimization problem provides the ability for scientists to use deep learning for scientific discovery, especially as an in situ data analysis tool for scientific simulations or data collection.

For more info: