Meta-Learning Dynamics Forecasting Using Task Inference

Explore meta-learning dynamics forecasting through task inference in this 53-minute lecture presented by Dr. Rose Yu at USC Information Sciences Institute. Delve into the challenges of generalization in deep learning models for dynamics forecasting and discover the innovative DyAd model-based meta-learning method. Learn how DyAd partitions heterogeneous domains into different tasks, utilizing an encoder for task inference and a forecaster for shared dynamics learning. Examine the theoretical foundations of generalization error and its relationship to task relatedness and domain differences. Gain insights into the model's performance on turbulent flow and real-world ocean data forecasting tasks, and understand its advantages over state-of-the-art approaches. Benefit from Dr. Yu's expertise in large-scale spatiotemporal data analysis and physics-guided AI as she discusses climate dynamics, computational challenges, and the acceleration of turbulence simulations.

Intro
Climate Dynamics
Computational Challenge
Dynamic Forecasting
Accelerating Turbulence Simulation
Generalization Challenge
Meta-Learning Dynamics
Related Work
Adaptation
Benefit of Multi-Task
Generalization Error
Experiments
Performance Comparison
Turbulent Flow Prediction
Ablation Study
Conclusion
Acknowledgment