Bayesian inversion (I):

Similar documents
Introduction to Systems Analysis and Decision Making Prepared by: Jakub Tomczak

DRAGON ADVANCED TRAINING COURSE IN ATMOSPHERE REMOTE SENSING. Inversion basics. Erkki Kyrölä Finnish Meteorological Institute

ECE295, Data Assimila0on and Inverse Problems, Spring 2015

USING GEOSTATISTICS TO DESCRIBE COMPLEX A PRIORI INFORMATION FOR INVERSE PROBLEMS THOMAS M. HANSEN 1,2, KLAUS MOSEGAARD 2 and KNUD S.

Miscellany : Long Run Behavior of Bayesian Methods; Bayesian Experimental Design (Lecture 4)

9/12/17. Types of learning. Modeling data. Supervised learning: Classification. Supervised learning: Regression. Unsupervised learning: Clustering

Lecture 4: Probabilistic Learning. Estimation Theory. Classification with Probability Distributions

Bayesian Models in Machine Learning

Lecture : Probabilistic Machine Learning

ENGG2430A-Homework 2

When one of things that you don t know is the number of things that you don t know

Consider the joint probability, P(x,y), shown as the contours in the figure above. P(x) is given by the integral of P(x,y) over all values of y.

Naïve Bayes classification

Lecture 4: Probabilistic Learning

Learning Gaussian Process Models from Uncertain Data

Machine Learning 4771

Course on Inverse Problems

Inverse problems and uncertainty quantification in remote sensing

Naïve Bayes classification. p ij 11/15/16. Probability theory. Probability theory. Probability theory. X P (X = x i )=1 i. Marginal Probability

Perhaps the simplest way of modeling two (discrete) random variables is by means of a joint PMF, defined as follows.

Parameter estimation! and! forecasting! Cristiano Porciani! AIfA, Uni-Bonn!

Statistical Learning Reading Assignments

Bayesian Methods and Uncertainty Quantification for Nonlinear Inverse Problems

Some Concepts of Probability (Review) Volker Tresp Summer 2018

Multiple Scenario Inversion of Reflection Seismic Prestack Data

Bayesian Approach 2. CSC412 Probabilistic Learning & Reasoning

Parametric Models. Dr. Shuang LIANG. School of Software Engineering TongJi University Fall, 2012

The local dark matter halo density. Riccardo Catena. Institut für Theoretische Physik, Heidelberg

Introduction to Probabilistic Machine Learning

Machine Learning Lecture 2

Regularized Regression A Bayesian point of view

A novel determination of the local dark matter density. Riccardo Catena. Institut für Theoretische Physik, Heidelberg

Posterior Regularization

Learning Bayesian network : Given structure and completely observed data

Nonlinear Bayesian joint inversion of seismic reflection

CSci 8980: Advanced Topics in Graphical Models Gaussian Processes

Machine Learning Srihari. Probability Theory. Sargur N. Srihari

Machine Learning Lecture 2

Density Estimation. Seungjin Choi

Nonparameteric Regression:

Bayesian Inference in Astronomy & Astrophysics A Short Course

Day 5: Generative models, structured classification

Lecture 1: Bayesian Framework Basics

Encoding or decoding

Probabilistic Reasoning in Deep Learning

Strong Lens Modeling (II): Statistical Methods

Reliability Monitoring Using Log Gaussian Process Regression

Bayesian Learning. CSL603 - Fall 2017 Narayanan C Krishnan

Learning with Probabilities

Uncertainty quantification for Wavefield Reconstruction Inversion

Bayesian Learning. HT2015: SC4 Statistical Data Mining and Machine Learning. Maximum Likelihood Principle. The Bayesian Learning Framework

Probability. Raul Queiroz Feitosa

Math-Stat-491-Fall2014-Notes-I

STAT 499/962 Topics in Statistics Bayesian Inference and Decision Theory Jan 2018, Handout 01

CS Lecture 18. Topic Models and LDA

PROBABILITY AND INFERENCE

Should all Machine Learning be Bayesian? Should all Bayesian models be non-parametric?

Probabilistic classification CE-717: Machine Learning Sharif University of Technology. M. Soleymani Fall 2016

Intelligent Systems Statistical Machine Learning

DS-GA 1002 Lecture notes 11 Fall Bayesian statistics

Density Estimation: ML, MAP, Bayesian estimation

Gaussian processes. Chuong B. Do (updated by Honglak Lee) November 22, 2008

Review: Probability. BM1: Advanced Natural Language Processing. University of Potsdam. Tatjana Scheffler

Machine Learning Lecture 2

{ p if x = 1 1 p if x = 0

Probabilistic Graphical Models

Signal detection theory

Probability and Information Theory. Sargur N. Srihari

Statistics and Data Analysis

Graphical Models and Kernel Methods

Machine Learning. Probability Theory. WS2013/2014 Dmitrij Schlesinger, Carsten Rother

Lecture 5. G. Cowan Lectures on Statistical Data Analysis Lecture 5 page 1

Conditional Random Field

COS513 LECTURE 8 STATISTICAL CONCEPTS

Estimation of Cole-Cole parameters from time-domain electromagnetic data Laurens Beran and Douglas Oldenburg, University of British Columbia.

The Bayesian approach to inverse problems

Gaussian Processes (10/16/13)

Inference and estimation in probabilistic time series models

Introduction to Bayesian inference

Aarti Singh. Lecture 2, January 13, Reading: Bishop: Chap 1,2. Slides courtesy: Eric Xing, Andrew Moore, Tom Mitchell

Grundlagen der Künstlichen Intelligenz

Winter 2019 Math 106 Topics in Applied Mathematics. Lecture 9: Markov Chain Monte Carlo

Introduction to Bayesian Learning

Intelligent Systems Statistical Machine Learning

Probabilistic and Bayesian Machine Learning

(3) Review of Probability. ST440/540: Applied Bayesian Statistics

New Bayesian methods for model comparison

PATTERN RECOGNITION AND MACHINE LEARNING

Recall from last time. Lecture 3: Conditional independence and graph structure. Example: A Bayesian (belief) network.

Joint Gaussian Graphical Model Review Series I

4 Pairs of Random Variables

BXA robust parameter estimation & practical model comparison. Johannes Buchner. FONDECYT fellow

The Bayes classifier

Advanced Statistical Methods. Lecture 6

A Frequentist Assessment of Bayesian Inclusion Probabilities

Bayesian Learning (II)

Bayesian Inference and MCMC

ORIGINS OF STOCHASTIC PROGRAMMING

Decision theory. 1 We may also consider randomized decision rules, where δ maps observed data D to a probability distribution over

Treatment and analysis of data Applied statistics Lecture 6: Bayesian estimation

Transcription:

Bayesian inversion (I): Advanced topics Wojciech Dȩbski Instytut Geofizyki PAN debski@igf.edu.pl Wydział Fizyki UW, 13.10.2004

Wydział Fizyki UW Warszawa, 13.10.2004 (1) Plan of the talk I. INTRODUCTION forward/inverse problems inversion as a parameter estimation task direct/indirect measurements II. PROBABILISTIC APPROACH inversion as an inference process description of information - probability experimental, theoretical, and apriori information Building A posteriori pdf Bayesian rule - a posteriori probability Tarantola s (probabilistic) approach A posteriori pdf - examples II. GEOPHYSICAL EXAMPLES

Forward and inverse problems forward modelling Wydział Fizyki UW Warszawa, 13.10.2004 (2)

Forward and inverse problems inversion Wydział Fizyki UW Warszawa, 13.10.2004 (3)

Forward and inverse problems parameter estimation Wydział Fizyki UW Warszawa, 13.10.2004 (4)

Wydział Fizyki UW Warszawa, 13.10.2004 (5) Model and Data Parameters Physical System: Model parameters: p 1, p 2, p K Predicted (Observed) data: m = (m 1, m 2, m M ) d = (d 1, d 2, d N ) Forward modelling: d = f(m)

Wydział Fizyki UW Warszawa, 13.10.2004 (6) Linear Inverse Problem Algebraic approach d = G m Naive inversion: m est = G 1 d Algebraic method: G T d = G T G m G T G = G T G + λi m est = (G T G + λi) 1 d

Wydział Fizyki UW Warszawa, 13.10.2004 (7) Inverse Problem Optimization approach Searching the model space for the model which best fits the data Least squares method: (d obs f(m)) T (d obs f(m)) = min More generally: (d obs f(m)) D + m m apr M = min

Wydział Fizyki UW Warszawa, 13.10.2004 (8) Direct / Indirect measurements Counts: events No. of particles quantified parameters Counts in a scale unit: mass luminosity temperature Quantities unmeasureable directly: mas of a earth, star, galaxy temperature distribution in the Earth, the sun, etc. mass of elementary particles...

Direct & Indirect measurements Wydział Fizyki UW Warszawa, 13.10.2004 (9)

Direct & Indirect measurements Wydział Fizyki UW Warszawa, 13.10.2004 (10)

Indirect measurement - Example Wydział Fizyki UW Warszawa, 13.10.2004 (11)

Inversion Back projection Wydział Fizyki UW Warszawa, 13.10.2004 (12)

Wydział Fizyki UW Warszawa, 13.10.2004 (13) Inversion optimization S(ρ) = V obs V (ρ) = min

Inversion as Joining Information (Inference) Wydział Fizyki UW Warszawa, 13.10.2004 (14)

Description of information (a) (data) Wydział Fizyki UW Warszawa, 13.10.2004 (15)

Description of information (a) (probability) Wydział Fizyki UW Warszawa, 13.10.2004 (16)

Description of information (b) (data) Wydział Fizyki UW Warszawa, 13.10.2004 (17)

Description of information (b) (probability) Wydział Fizyki UW Warszawa, 13.10.2004 (18)

Description of information (b) (data) Wydział Fizyki UW Warszawa, 13.10.2004 (19)

Description of information (b) (probability) Wydział Fizyki UW Warszawa, 13.10.2004 (20)

Frequentists interpretation of probability Wydział Fizyki UW Warszawa, 13.10.2004 (21)

Bayesian interpretation of probability Wydział Fizyki UW Warszawa, 13.10.2004 (22)

Joining Information (Inference) Wydział Fizyki UW Warszawa, 13.10.2004 (23)

Wydział Fizyki UW Warszawa, 13.10.2004 (24) Experimental information Experimental uncertainties: ɛ: random = p(ɛ) d obs = d true + ɛ P ([ɛ δ/2, ɛ + δ/2]) P(d = d true ) = p ɛ (d d obs ) But how to find p(ɛ)???

Wydział Fizyki UW Warszawa, 13.10.2004 (25) Experimental information Experimental uncertainties statistic: postulate: p(ɛ) = k e ɛ2 2σ 2 estimated by repeating measurement P ([x δ/2, x + δ/2]) = N i N

A priori information Wydział Fizyki UW Warszawa, 13.10.2004 (26)

Wydział Fizyki UW Warszawa, 13.10.2004 (27) Theoretical information Theoretical information: correlation d = G(m) = p(d = G(m) m) G may be known approximatelly: G(m) = G o m + G 1 m 2 +... Exact theory: p(d m) = δ(d G(m))

Theoretical information Wydział Fizyki UW Warszawa, 13.10.2004 (28)

Theoretical information Wydział Fizyki UW Warszawa, 13.10.2004 (29)

Theoretical information Wydział Fizyki UW Warszawa, 13.10.2004 (30)

Wydział Fizyki UW Warszawa, 13.10.2004 (31) Multi-dimensional pdf p(x, y) marginal pdf p x (x) = Y marginal pdf p y (y) = X p(x, y)dy p(x, y)dx conditional pdf: p x y (x y) = p(x, y)/p y (y) conditional pdf: p y x (y x) = p(x, y)/p x (x)

Wydział Fizyki UW Warszawa, 13.10.2004 (32) Multi-dimensional pdf p(x, y) = p x y (x y)p y (y) p(x, y) = p y x (y x)p x (x) p x y (x y)p y (y) = p y x (y x)p x (x)

Wydział Fizyki UW Warszawa, 13.10.2004 (33) Multi-dimensional pdf p m (m d) = p d m(d m)p m (m) p d (d)

Wydział Fizyki UW Warszawa, 13.10.2004 (34) Joining information 1. observation: p(d) 2. theory: q(m, d) 3. a priori f(m, d) p q(x) = p(x) q(x) µ(x)

Conditional vs. joint pdf Wydział Fizyki UW Warszawa, 13.10.2004 (35)

Wydział Fizyki UW Warszawa, 13.10.2004 (36) A posteriori pdf A posteriori pdf: σ(m, d) p(d) q(m, d) f M (m)f D (m) describes all information on d, m. σ m (m) = σ(m, d)dd σ d (d) = D σ(m, d)dm M

Wydział Fizyki UW Warszawa, 13.10.2004 (37) A posteriori pdf σ m (m) = f M (m) L(m, d obs )

Wydział Fizyki UW Warszawa, 13.10.2004 (38) A posteriori pdf A posteriori pdf σ(m, d): always exists is unique describes all information is the solution of an inverse problem

Wydział Fizyki UW Warszawa, 13.10.2004 (39) Exact theory σ(m) = f M (m) D p(d)q(m, d)dd q(d, m) = δ(d G(m)) σ(m) = f M (m) p(d G(m)) σ(m) = f M (m) exp ( (d G(m)) T C 1 (d G(m)) )

Exact theory: linear case Wydział Fizyki UW Warszawa, 13.10.2004 (40)

Exact theory: non-linear case Wydział Fizyki UW Warszawa, 13.10.2004 (41)

Exact data: uncertain theory Wydział Fizyki UW Warszawa, 13.10.2004 (42)

Non-resolved parameters Wydział Fizyki UW Warszawa, 13.10.2004 (43)

Wydział Fizyki UW Warszawa, 13.10.2004 (44) Summary Parameter estimation Error analysis Resolution analysis Experimental planning Model discrimination Others (non-parametric) inference

Wydział Fizyki UW Warszawa, 13.10.2004 (45) Inversion Algorithms- Summary Method Advantages Limitations Algebraic (LSQR) - Simplicity - Only linear problems m ml = (G T G + γi) 1 G T d obs - Large scale problems - Lack of robustness Optimization - Simplicity - Difficult error estimation G(m) d obs + λ m) m a = min - Fully nonlinear Bayesian - Fully nonlinear - More complex theory σ(m) = f(m)l(m, d obs ) - Full error handling - Requires efficient sampler