Abstract: In this talk, we outline recent additions to DeepHyper, a scalable, multipurpose automatic machine learning framework being developed at Argonne. Previously, DeepHyper was successfully deployed for large-scale neural architecture searches of fully connected networks to model drug responses for the CANDLE benchmark datasets. In this talk, we focus on searching for recurrent neural architectures to predict the spectral content of geophysical datasets and thereby obtain effective surrogate models that provide forecasts very efficiently. Subsequently, we evaluate their accuracy using NOAA’s 30-year sea-surface temperature dataset as well as NASA’s DayMet repository for maximum temperatures over North America.
Our assessments show that the architectures obtained by DeepHyper are seen to compare favorably to traditional time-series modeling methods. Also, scaling analyses of the architecture search algorithms demonstrate effective node utilization up to 512 nodes of Theta. Finally, we will conclude this talk by providing a glimpse of upcoming additions to DeepHyper such as the ability to search for convolutional neural architectures for image-based datasets.