Host: Jiapeng Gao
Title: Title: Machine Learning Across Scales: From Hidden Exoplanets to Efficient Model Training
Abstract:
Non-transiting planets can remain hidden while leaving dynamical signals in the transit timing variations (TTVs) of observed planets. DeepTTV provides a machine learning (ML) approach to this inverse problem, using N-body simulations to generate training data for a neural network that combines recurrent and Transformer architectures to infer the mass and orbital properties of unseen planetary companions. In the first part of this talk, I will discuss DeepTTV. In the second part, I will introduce my current work in industry on improving the efficiency of large-scale ML training, including directions such as transfer learning and improving computational resource utilization.
]]>Abstract:
Non-transiting planets can remain hidden while leaving dynamical signals in the transit timing variations (TTVs) of observed planets. DeepTTV provides a machine learning (ML) approach to this inverse problem, using N-body simulations to generate training data for a neural network that combines recurrent and Transformer architectures to infer the mass and orbital properties of unseen planetary companions. In the first part of this talk, I will discuss DeepTTV. In the second part, I will introduce my current work in industry on improving the efficiency of large-scale ML training, including directions such as transfer learning and improving computational resource utilization.
]]>