"Augmenting Incomplete Physical Models with Deep Networks for Complex Dynamic Forecasting"
This talk addresses the problem of combining Machine Learning (ML) and Model-Based (MB) approaches for complex dynamic forecasting. We especially tackle the situation where a prior physical knowledge formalized through ODE/PDE is available, but is incomplete in the sense that it cannot fully describe the observed dynamics.As a first contribution, I introduce the PhyDNet model for predicting the whole content of future images in unsupervised videos forecasting. A latent space is learned such that that a simple linear PDE dynamics, e.g. corresponding to object motions, is applicable. PhyDNet is a two-branch recurrent neural network: one branch is responsible for modeling the physical dynamics in the latent space while the other branch captures the complementary information required for accurate prediction. Experiments on various video datasets show the relevance of the approach. To study more in depth the cooperation between ML and MB models, I concentrate on the decomposition between both components. I introduce a principled learning framework, called APHYNITY which minimizes the norm of the data-driven complement under the constraint of perfect prediction of the augmented model. I provide a theoretical analysis of the decomposition and show that we can ensure existence and uniqueness decomposition guarantees, under mild conditions. I will present several applications where ODE/PDE dynamics can be augmented with APHYNITY to ensures better forecasting and parameter identification performances than MB or ML models alone, and that competing MB/ML hybrid methods.
- "Disentangling Physical Dynamics from Unknown Factors for Unsupervised Video Prediction." Vincent Le Guen, Nicolas Thome, CVPR'20.
- "A Deep Physical Model for Solar Irradiance Forecasting with Fisheye Images." Vincent Le Guen, Nicolas Thome, CVPR’20 OMNI-CV workshop.
- "Augmenting Physical Models with Deep Networks for Complex Dynamics Forecasting." Yuan Yin, Vincent Le Guen, Jérémie Dona, Ibrahim Ayed, Emmanuel de Bézenac, Nicolas Thome, Patrick Gallinari, ICLR’21 (oral presentation).
Nicolas Thome is a full professor at Conservatoire Nationnal des Arts et Métiers (Cnam Paris). His research interests include machine learning and deep learning for understanding low-level signals, e.g. vision, time series, acoustics, etc. His current application domains are essentially targeted towards healthcare, autonomous driving and physics. He is involved in several French, European and international collaborative research projects on artificial intelligence and deep learning.
*If you are a Stanford Affiliate outside of Energy Resources Engineering department and would like to attend, please contact Emily Gwynn (email@example.com) for the Zoom Meeting link and passcode.
*If you are in Energy Resources Engineering department, you will receive an announcement from Emily Gwynn (firstname.lastname@example.org) with the Zoom Meeting link and passcode.