Accelerating Machine Learning Model Design with Automated Pipeline and HPC Resources
Machine learning (ML) tasks such as feature engineering, model selection, training, and tuning are computationally expensive and time-consuming, especially on limited hardware. In response to this dilemma, an Automated Machine Learning Pipeline (AMPL) was developed to streamline the design of ML models on large-scale datasets across multiple domains. Coupled with the utilization of DoD high-performance computing resources, AMPL accelerates model development and training, making advanced ML approaches more feasible. From data pre-processing to model evaluation, AMPL provides the framework for automating these tasks while lowering the barrier to HPC by obscuring the underlying infrastructure. In addition to streamlining the ML development pipeline with task automation, AMPL also allows users to refine ML models and improve performance by integrating advanced methods. AMPL has been successfully applied to problems within military domains, such as evaluating rotor blade performance and estimating kinetic energy reduction for terminal ballistics.
IMPACT
Accomplishment: Successfully developed an automated, multi-domain machine learning pipeline tool, designing and training models for problems within military domains; Result: Reduced the timeline to design and train an initial model from months to days – Enhanced model design capabilities and reduced computational costs.
PRESENTER
Thompson, Brianna
brianna.d.thompson@erdc.dren.mil
601-634-7193
U.S. Army Engineer Research and Development Center
CO-AUTHOR(S)
Ross, James
james.e.ross@erdc.dren.mil
Kubiak, Lisa
lisa.a.kubiak@erdc.dren.mil
CATEGORY
AI/ML for HPC
SYSTEM(S) USED
Carpenter, Narwhal