LANS Informal Seminar: Prasanna Balaprakash

Seminar Title: Scalable Reinforcement-Learning-Based Neural Architecture Search for Scientific and Engineering Applications
Speaker: Prasanna Balaprakash, Computer Scientist (MCS & LCF, ANL)

Date/Time: 2019-02-06 10:30
Location: Bldg. 240, Rm. 1404-05

Description:
The success of deep learning in machine learning applications has encouraged
the scientific and engineering community to develop deep-learning-based
predictive models for a wide a range of applications. Designing a deep neural
network (DNN) architecture for a particular modeling task, however, requires
significant architecture engineering by a deep learning expert. While several
recent works discuss automating the process of the neural architecture search
(NAS), they have focused mainly on the traditional machine learning tasks of
image and text classification. In this talk, we will present a scalable NAS
approach to automatically generate DNN models for predictive modeling in
science and engineering applications. We will discuss a recurrent
neural-network-based architecture generator that produces a multilayered
perceptron with skip connections. We leverage a manager-worker-based
distributed reinforcement-learning approach using proximal policy optimization
method to iteratively improve the generated DNN architectures. We demonstrate
the effectiveness of the proposed NAS approach for multivariate and multioutput
regression problems on diverse applications. The generated architectures obtain
high accuracy while maintaining significantly fewer parameters and achieve 70%
to 80% node utilization on 256 to 1,024 nodes of Theta supercomputer at ALCF.