CILVR SEMINAR: Large learning rate in matrix recovery problems | Data-driven multiscale modeling of subgrid parameterizations in climate models
Date:
Thursday, October 5, 2023,
11AM
Location:
60FA
,
Room
7th floor common area
Speaker:
Lei Chen and Karl Otness
Subgrid parameterizations, which represent physical processes occurring below the resolution of current climate models, are an important component in producing accurate, long-term predictions for the climate. In this talk Karl will present this research problem and discuss an ongoing project investigating a multiscale, deep learning-based approach to this task. Lei will introduce some of his ongoing research on optimization instability and implicit bias.
Notes:
Abstract for Lei’s talk:
Lei will introduce some of his ongoing research on optimization instability and implicit bias. He will introduce the setting of Edge of Stability, where the sharpness of loss landscape is beyond the stability threshold. Then he will talk about the implicit bias of large learning rates in matrix factorization and matrix sensing.
Lei will introduce some of his ongoing research on optimization instability and implicit bias. He will introduce the setting of Edge of Stability, where the sharpness of loss landscape is beyond the stability threshold. Then he will talk about the implicit bias of large learning rates in matrix factorization and matrix sensing.
Bio:
Lei Chen is a fourth-year CS PhD student advised by Prof. Joan Bruna. His research interests include theoretical questions of graph neural networks. More recently he has been working on optimization in neural networks, especially with large learning rates.
Lei Chen is a fourth-year CS PhD student advised by Prof. Joan Bruna. His research interests include theoretical questions of graph neural networks. More recently he has been working on optimization in neural networks, especially with large learning rates.
Abstract for Karl’s talk:
Subgrid parameterizations, which represent physical processes occurring below the resolution of current climate models, are an important component in producing accurate, long-term predictions for the climate. Recent works have proposed a variety of machine learning approaches to modeling these unresolved dynamics. In this talk I will present this research problem and discuss an ongoing project investigating a multiscale, deep learning-based approach to this task.
Subgrid parameterizations, which represent physical processes occurring below the resolution of current climate models, are an important component in producing accurate, long-term predictions for the climate. Recent works have proposed a variety of machine learning approaches to modeling these unresolved dynamics. In this talk I will present this research problem and discuss an ongoing project investigating a multiscale, deep learning-based approach to this task.
Bio:
Karl Otness is a PhD student at NYU advised by professors Joan Bruna and Benjamin Peherstorfer. His research focuses on applications of machine learning to modeling and simulation problems.
Karl Otness is a PhD student at NYU advised by professors Joan Bruna and Benjamin Peherstorfer. His research focuses on applications of machine learning to modeling and simulation problems.