We propose to use evolutionary optimization methods on the Titan Computer to design optimal convolutional neural networks (CNN) and spiking neural networks (SNN) to create a better and faster method to analyze scientific data. There are currently no existing methods for combining much less optimizing CNNs and SNNs. However, given the GPU intensity of CNNs, the CPU intensity of SNNs, and the success we have had using evolutionary optimization (EO) in neural network design, Titan is the ideal platform for this hybrid CPU/GPU optimization. We expect to use the entire machine to create a new class of machine learning capability and the ability to quickly put these methods into the hands of scientist. We have scaled several design methods for both convolutional and spiking neural networks using 100% of Titan with great success. Based on this prior success, the proposed work would enable us to quickly create classification models for scientific data based on a spatial snapshot in time, and looking at the temporal aspects of the data, with a vision of helping guide a scientist to new discoveries and experiments.
|Source||Hours||Start Date||End Date|
|DOE ALCC PROGRAM||200,000||2019-07-01||2020-06-30|
|DOE ALCC PROGRAM||5,000||2019-07-01||2020-06-30|
|DOE ALCC PROGRAM||25,000,000||2018-07-01||2019-07-31|
|DOE ALCC PROGRAM||5,000||2018-07-01||2019-06-28|
|DOE ALCC PROGRAM||2,350,000||2018-07-01||2019-06-28|