Supervisor: Justin M. Wozniak (Argonne National Laboratory)
Abstract: As deep learning continues to expand into new areas of application, the demand for efficient use of our HPC resources increase. For new problem domains, new model architectures are developed through a neural architecture search (NAS), which consist of iteratively training many neural networks. To combat the computational waste and maximize compute efficiency for NAS, we demonstrate that the use of genetic algorithms with speciation can be used to both shorten training time and increase accuracy at each iteration.
ACM-SRC Semi-Finalist: no
Poster Summary: PDF
Back to Poster Archive Listing