Scribd
Upload
57 views

Physics Aware Machine Learning

1) The document discusses using machine learning techniques like convolutional neural networks and long short-term memory networks for tasks like spatio-temporal image classification and variable prediction in domains like remote sensing. 2) It also discusses developing "physics-aware machine learning" by incorporating physical constraints and relationships into models or using hybrid modeling frameworks. This includes emulating complex simulation codes and encoding differential equations. 3) Understanding models and their interpretations is emphasized over just fitting data, through techniques like feature selection, neuron visualization, and discovering causal relationships.

Uploaded by

Santonu G
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
57 views

Physics Aware Machine Learning

1) The document discusses using machine learning techniques like convolutional neural networks and long short-term memory networks for tasks like spatio-temporal image classification and variable prediction in domains like remote sensing. 2) It also discusses developing "physics-aware machine learning" by incorporating physical constraints and relationships into models or using hybrid modeling frameworks. This includes emulating complex simulation codes and encoding differential equations. 3) Understanding models and their interpretations is emphasized over just fitting data, through techniques like feature selection, neuron visualization, and discovering causal relationships.

Uploaded by

Santonu G
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 17

Physics-aware and Explainable

Machine Learning
Gustau Camps-Valls, Luis Gómez-Chova, Daniel Svendsen
Diego Bueso, Luca Martino, Adrian Pérez-Suay
María Piles, Valero Laparra, Ana B. Ruescas

Image Processing Lab (IPL)


Universitat de València -– http://isp.uv.es
Standard machine learning

● Supervised learning
– Classification
– Regression & model inversion
– Anomaly/target detection
● Unsupervised learning
– Density estimation
– Dimensionality reduction
– Clustering
Standard supervised machine learning

F(X) = y
● X: observations, independent covariates
● Y: target, dependent variable
● F: machine learning model (nonlinear, nonparametric, flexible, learned from data)
#1 – Spatio-temporal image classification
● Convolutional neural nets (CNN): hierarchical structure exploits spatial relations
● Long short-term memory (LSTM): recurrent network that accounts for memory/dynamics

“A Deep Network Approach to Multitemporal Cloud Detection”


Tuia, Perez-Suay and Camps-Valls, IEEE IGARSS 2018, http://isp.uv.es/code/landmarks.html
#2- Spatio-temporal variable prediction

● STA is common place in climate informatics, neuroscience, video processing, NLP, ...
● Current approaches
– CNN + LSTM
– Space-time (deep) GPs
– Combine CNN+GPs

● Many applications in Remote Sensing


– Parameter retrieval / estimation
– Time series gap filling
– Sensor fusion / Data assimilation

“ A Survey on Gaussian Processes for Earth Observation Data Analysis” “ Statistical Retrieval of Atmospheric Pro-les with Deep Convolutional Neural Networks”,
Camps-Valls et al. IEEE Geoscience and Remote Sensing Magazine 2016 Malmgren-Hansen, D. and Laparra, V. and Camps-Valls, G., IEEE TGARS, 2018
Physics-aware machine learning

F(X, )=y
#1 - Physics-driven ML: constrained optimization
● ML that respects laws of physics (e.g. consistency model-data, convection, advection, mass and energy conservation)
Minimize model violations

“Theory-guided Data Science”, Karpatne, A. et al. IEEE Trans. Know. Data Eng., 2017.

Fair ML

“Fair Kernel Learning” Perez-Suay, Laparra, Gomez-Chova, Camps-Valls, G. et al. ECML, 2017.
Joint Model-Data ML

“Joint Gaussian Processes for Biophysical Parameter Retrieval” Svendsen, Martino, Camps-Valls, IEEE TGARS 2018
“Physics-aware Gaussian processes in remote sensing” Camps-Valls, G. et al. Applied Soft Computing, 2018.
#2 - Physics-driven ML: hybrid modeling framework
● ML that learns laws of physics (e.g. consistency model-data, convection, advection, mass and energy conservation)

A: “Physisizing” a deep learning B: A motion field is learned with a


architecture by adding one or several convolutional-deconvolutional net, and
physical layers after the multilayer the motion field is further processed
neural network with a physical model

“Deep learning and process understanding for “Deep Learning for Physical Processes:
data-driven Earth System Science” Incorporating Prior Scienti-c Knowledge”.
Reichstein, Camps-Valls et al. Nature, 2018. de Bezenac, Pajot, & Gallinari, arXiv:1711.07970 (2017).
#3- Physics-driven ML: emulation of complex codes
● GP Emulation = Mathematical tractability + Global sensitivity analysis + Speed

“Emulation of Leaf, Canopy and Atmosphere Radiative Transfer Models for Fast Global Sensitivity Analysis”,
Verrelst, Camps-Valls et al Remote Sensing of Environment, 2016
“Emulation as an accurate alternative to interpolation in sampling radiative transfer codes”,
Vicent and Camps-Valls, IEEE Journal Sel. Topics Rem. Sens, Apps. 2018
#3- Physics-driven ML: emulation of superparam. models
● NN Emulation of a superparameterized model: no virtual error, 10x faster

“Deep learning to represent subgrid processes in climate models” Rasp, Pritchard, Pierre Gentine, PNAS 2018
#4- Physics-driven ML: encoding and learning ODE/PDEs
● Who needs Lorenz? ● Who needs Navier Stokes?

● Who needs Schrödinger?

“Discovering governing equations from data by sparse identi-cation of nonlinear dynamical “Deep Hidden Physics Models: Deep Learning of Nonlinear Partial Di7erential Equations”
systems” Brunton, Proctor, Kutz, PNAS 2016 Raissi, JMLR 2018
Understanding is more important than fitting

F(X,y) =
#1- Feature selection & ranking

● Filters & wrappers

“Remote Sensing Feature Selection by Kernel Dependence Estimation”, Camps-Valls, G. Mooij, JM. Schölkopf, IEEE-GRSL, 2010.
“A guided hybrid genetic algorithm for feature selection with expensive cost functions”, M. Jung, J. Zscheischler, Procedia, 2013.
#2- Neuron and bases visualization

● What did the network learn?


● How do bases change in time, with real/simulations/together, under extremes?

“Visualizing and Understanding Convolutional Networks”, Zeriler, et al 2013


“Processing of Extremely high resolution LiDAR and optical data”, Campos-Taberner, Camps-Valls et al, 2016
“DeeplyOut: What did your network learn under anomalies and adaptation? ,” Camps-Valls et al, JMLR (2019)
#3- Graphical models and causality
● Causality discovery learns cause and effects relations from data
● What for? Hypothesis testing, model-data comparison, causes of extreme impacts

“Inferring causation from time series with perspectives in Earth system sciences”, Runge, Bathiany, Bollt, Camps-Valls, et al. Nat Comm (submitted), 2018.
“Causal Inference in Geoscience and Remote Sensing from Observational Data,” Pérez-Suay and Camps-Valls, IEEE Trans. Geosc. Rem. Sens, 2018
“CauseMe: An online system for benchmarking causal inference methods,” Muñoz-Marí, Mateo, Runge, Camps-Valls. In preparation (2019). CauseMe: http://causeme.uv.es
Conclusions
Conclusions

● Machine learning in EO and climate


○ Many techniques ready to use
○ Huge community, exciting tools
● Solid mathematical framework
○ Multivariate data
○ Multisource data
○ Structured spatio-temporal relations
○ Nonlinear feature relations
○ Fitting & understanding
● Risks: addictive, overfitting, overlooking
● Remedies: Physics-driven ML, Explainable AI, Causality

You might also like