Statistical approach for dictionary learning

Similar documents
Machine Learning for Signal Processing Sparse and Overcomplete Representations. Bhiksha Raj (slides from Sourish Chaudhuri) Oct 22, 2013

Machine Learning for Signal Processing Sparse and Overcomplete Representations

Bayesian Paradigm. Maximum A Posteriori Estimation

Introduction to Machine Learning CMU-10701

an introduction to bayesian inference

A graph contains a set of nodes (vertices) connected by links (edges or arcs)

STA 414/2104: Machine Learning

STA 4273H: Statistical Machine Learning

Machine Learning Techniques for Computer Vision

Hidden Markov Models

Restricted Boltzmann Machines for Collaborative Filtering

Recent developments on sparse representation

Expectation Maximization

Review: Learning Bimodal Structures in Audio-Visual Data

STA 4273H: Statistical Machine Learning

Lecture 14. Clustering, K-means, and EM

Bayesian Networks: Construction, Inference, Learning and Causal Interpretation. Volker Tresp Summer 2016

Hidden Markov Models. By Parisa Abedi. Slides courtesy: Eric Xing

Variational Inference (11/04/13)

Julesz s Quest for texture perception

Recent Advances in Bayesian Inference Techniques

CPSC 540: Machine Learning

The connection of dropout and Bayesian statistics

EUSIPCO

Hidden Markov Models. Aarti Singh Slides courtesy: Eric Xing. Machine Learning / Nov 8, 2010

Rigorous Dynamics and Consistent Estimation in Arbitrarily Conditioned Linear Systems

A minimalist s exposition of EM

MCMC: Markov Chain Monte Carlo

Statistical learning. Chapter 20, Sections 1 4 1

Single-channel source separation using non-negative matrix factorization

Review. DS GA 1002 Statistical and Mathematical Models. Carlos Fernandez-Granda

Consistency of the maximum likelihood estimator for general hidden Markov models

Introduction to Restricted Boltzmann Machines

Introduction to Machine Learning CMU-10701

Lecture 16 Deep Neural Generative Models

Pattern Recognition and Machine Learning

CSE 150. Assignment 6 Summer Maximum likelihood estimation. Out: Thu Jul 14 Due: Tue Jul 19

State-Space Methods for Inferring Spike Trains from Calcium Imaging

U-Likelihood and U-Updating Algorithms: Statistical Inference in Latent Variable Models

Variational Methods in Bayesian Deconvolution

A tutorial on sparse modeling. Outline:

Machine learning - HT Maximum Likelihood

Pattern Recognition and Machine Learning. Bishop Chapter 11: Sampling Methods

Gaussian Process Approximations of Stochastic Differential Equations

Bayesian Networks: Construction, Inference, Learning and Causal Interpretation. Volker Tresp Summer 2014

Sparse Time-Frequency Transforms and Applications.

Computer Intensive Methods in Mathematical Statistics

Sparse & Redundant Signal Representation, and its Role in Image Processing

PATTERN RECOGNITION AND MACHINE LEARNING CHAPTER 2: PROBABILITY DISTRIBUTIONS

Markov Chains and Hidden Markov Models

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

Self-Organization by Optimizing Free-Energy

Hidden Markov Models Part 2: Algorithms

Expectation Maximization (EM)

13: Variational inference II

The Particle Filter. PD Dr. Rudolph Triebel Computer Vision Group. Machine Learning for Computer Vision

Graphical Models for Collaborative Filtering

Note Set 5: Hidden Markov Models

Week 3: The EM algorithm

STA414/2104 Statistical Methods for Machine Learning II

Dimensionality Reduction. CS57300 Data Mining Fall Instructor: Bruno Ribeiro

Modeling Data with Linear Combinations of Basis Functions. Read Chapter 3 in the text by Bishop

Study Notes on the Latent Dirichlet Allocation

Variational Autoencoder

Parametric Inference Maximum Likelihood Inference Exponential Families Expectation Maximization (EM) Bayesian Inference Statistical Decison Theory

Pattern Recognition and Machine Learning. Bishop Chapter 9: Mixture Models and EM

Basic Principles of Unsupervised and Unsupervised

Markov Networks. l Like Bayes Nets. l Graph model that describes joint probability distribution using tables (AKA potentials)

Applications of Hidden Markov Models

5 Mutual Information and Channel Capacity

Latent Dirichlet Alloca/on

A = {(x, u) : 0 u f(x)},

Sparse Linear Models (10/7/13)

Machine Learning and Bayesian Inference. Unsupervised learning. Can we find regularity in data without the aid of labels?

CS839: Probabilistic Graphical Models. Lecture 7: Learning Fully Observed BNs. Theo Rekatsinas

Expectation Propagation Algorithm

Energy Based Models. Stefano Ermon, Aditya Grover. Stanford University. Lecture 13

Index. Santanu Pattanayak 2017 S. Pattanayak, Pro Deep Learning with TensorFlow,

Expectation Maximization

Data Preprocessing. Cluster Similarity

STA 4273H: Statistical Machine Learning

Latent Tree Approximation in Linear Model

Independent Component Analysis and Unsupervised Learning

Blind Equalization via Particle Filtering

Machine Learning & Data Mining Caltech CS/CNS/EE 155 Hidden Markov Models Last Updated: Feb 7th, 2017

Statistical learning. Chapter 20, Sections 1 3 1

Linear Models for Regression CS534

ECE521 lecture 4: 19 January Optimization, MLE, regularization

Probability Based Learning

Joint Factor Analysis for Speaker Verification

Independent Component Analysis. Contents

Tensor-Based Dictionary Learning for Multidimensional Sparse Recovery. Florian Römer and Giovanni Del Galdo

Heeyoul (Henry) Choi. Dept. of Computer Science Texas A&M University

Lecture 8 Learning Sequence Motif Models Using Expectation Maximization (EM) Colin Dewey February 14, 2008

Independent Component Analysis and Unsupervised Learning. Jen-Tzung Chien

Probabilistic and Bayesian Machine Learning

Learning Bayesian Networks

Auto-Encoding Variational Bayes

HMM: Parameter Estimation

Learning Sequence Motif Models Using Expectation Maximization (EM) and Gibbs Sampling

Transcription:

Statistical approach for dictionary learning Tieyong ZENG Joint work with Alain Trouvé Page 1

Introduction Redundant dictionary Coding, denoising, compression. Existing algorithms to generate dictionary K-SVD MOTIF Our object: find a dictionary more pertinent to a class of image translation-invariant Page 2

M. Elad and M. Aharon 2006 K-SVD results Page 3

MOTIF P. Jost, S. Lesage, P. Vandergheynst and R. Gribonval, 2006 Results of MOTIF: 19 generating functions learnt on natural images Page 4

Plan Generative model Bernoulli-Exponential Model (BEM) Bernoulli-Gaussian Model (BGM) Identifiably issues MCMC-EM for BEM Mean field method for BGM Experiments Conclusion Page 5

Generative model We denote: the discrete tore, support of images the spaces of images a family of local atoms generating the dictionary translation of the atoms over the plan We consider the following generative model : where Bernoulli variable, i.i.d, flag: active or disactive random coefficient, i.i.d, intensity white noise Goal: given, to learn Page 6

Two models: BEM and BGM The Bernoulli-Exponential Model (BEM) is of exponential distribution Parameter space The complete likelihood The Bernoulli-Gaussian Model (BGM) is of Gaussian distribution Parameter space Page 7

Identifiably problem We are interested in a statistical question: Can we distinguish two distributions on Y given by two different parameters? For both models: under weak conditions, YES! BEM Equivalent relation: if and modulo a permutation on the indices Prop. If are different, then the BEM is identifiable in BGM Similarly result holds. Page 8

Expectation-Maximization method Recall that our model is: B h,s,x h,s where are hidden variables. Usually we can use the EM method: E-step: specify a value for parameter, get the hidden variables by the conditional expectation M-step: given hidden variables, get a value for parameter by maximization the likelihood Page 9

From likelihood to MCMC The approach of maximum likelihood calculates The method of EM iterates where is the a posteriori distribution of (X, B) known Y with Z is a normalization factor and is the dominated measure. (E)-step, (M)-step We can approximate the a posteriori distribution by MCMC: we generate a Markov chain associated to Page 10

Pseudo-code of MCMC-EM Algorithm of MCMC-EM Page 11

Accept-rejection algorithm Suppose that we have a target distribution and an auxiliary distribution such that on the support of. Then we can simulate by the flowing algorithm: 1. GenerateX πandu U([0,1]) 2. AcceptY =X if 3. ReturnStep1ifrejected Page 12

The update of θ Now, let us look at the update of θ. Since, Calculating in a straightforward manner, we have, Page 13

How to update Φ In order to update the dictionary, we need solve: We first solve this problem without constraints, then project the elements on sphere by normalization. Denoting this is equivalent to: Page 14

Plan Generative model Bernoulli-Exponential Model (BEM) Bernoulli-Gaussian Model (BGM) Identifiably issues MCMC-EM for BEM Mean field method for BGM Experiments Conclusion Page 15

Mean field method for BGM is very difficult to calculate Mean field: we approximate by a product of distribution on and Considering the collection of distribution product on is of Bernoulli distribution is of distribution Then we take and such that where K is the Kullback-Leibler divergence: Page 16

The fixed point equation and update of the parameter For the Kullback-Leibler divergence, the optimal distribution satisfies a fixed point equation Then, the step of updating the parameter is given by In consequence, we have: Page 17

Numerical aspects Grids for fixed point equation Thresholding to get sparse elements Support compact where Initialization of parameters Matching Pursuit Page 18

Experiments for MCMC and mean field Training set Ideal dictionary MCMC Mean field Page 19

Experiment of mean field Example of Training set Real dictionary Mean field Page 20

Experiment of mean field Real dictionary Mean field Page 21

Mean field for natural image Page 22

Conclusion Generative model Bernoulli-Exponential Model (BEM) Bernoulli-Gaussian Model (BGM) Identifiably issues MCMC-EM for BEM Mean field method for BGM Experiments Future works Theoretical aspects, consistence Rigid atoms ==== Deformation atoms Natural images Page 23

Thank you! Page 24