Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd
Upload
9 views

Tutorial em

Uploaded by

minda14650
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
9 views

Tutorial em

Uploaded by

minda14650
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 57

Gaussian Mixture Models

and EM algorithm
Radek Danecek
Gaussian Mixture Model
• Unsupervised method
• Fit multimodal Gaussian distributions
Formal Definition
• The model is described as:

• The parameters of the model are:

• The training data is unlabeled – unsupervised setting

• Why not fit with MLE?


Optimization problem
• Model:

• Apply MLE:
• Maximize:

• Difficult, non convex optimization with constraints

• Use EM algorithm instead


EM Algorithm for GMMs
• Idea:
• Objective function:
• Split optimization of the objective into to parts

• Algorithm:
• Initialize model parameters (randomly):
• Iterate until convergence:
• E-step
• Assign cluster probabilities (“soft labels”) to each sample

• M-step
• Solve the MLE using the soft labels
Initialization
• Initialize model parameters (randomly)

• Uniform for cluster probabilities


• Centers
• Random
• K-means heuristics

• Covariances:
• Spherical, according to empirical variance
E-step
K
• For each data point and each
cluster , compute the probability
that belongs to Probabilities of point n
belonging to clusters 1…K
(given current model parameters) (sum up to 1)
N

Sum determines the


“weight” of cluster k

“soft labels”
E-step
• For each data point and each
cluster , compute the probability
that belongs to
(given current model parameters)

“soft labels”
M-step
• Now we have “soft labels” for the data -> fall back to supervised MLE

• Optimize the log likelihood:


• Instead of the original (difficult objective):
We optimize the following:

• Differentiate w.r.t.
M-step
K
• Update model parameters:

Probabilities of point n
belonging to clusters 1…K
(sum up to 1)
• Update prior for each cluster: N

Sum determines the


“weight” of cluster k
Normalized column-wise sum
are priors for clusters 1…K
M-step
• Update model parameters:

• Update mean and covariance of


each cluster
M-step
• Update model parameters:

• Update mean and covariance of


each cluster
EM Algorithm for GMMs
• Idea:
• Objective function:
• Split optimization of the objective into to parts
• Algorithm:
• Initialize model parameters (randomly):
• Iterate until convergence:
• E-step
• Assign cluster probabilities (“soft labels”) to each sample
• M-step
• Find optimal parameters given the soft labels
Overlapping clusters
Overlapping clusters
Unequal cluster size
Imbalanced cluster size
Sensitivity to Initialization
Sensitivity to Initialization
Sensitivity to Initialization
Sensitivity to Initialization
Sensitivity to Initialization
Sensitivity to initialization
Degenerate covariance
• The determinant of the covariance matrix tends to 0
Practical Example – Color segmentation
• Input: an image

• Can be thought of as a dataset of 3D (color) samples

• Run 3D GMM clustering over


Practical Example – Color segmentation

Input Image
Practical Example – Color segmentation

3 clusters
Practical Example – Color segmentation

4 clusters
Practical Example – Color segmentation

5 clusters
Practical Example – Color segmentation

7 clusters
Practical Example – Color segmentation

8 clusters
Practical Example – Color segmentation

9 clusters
Practical Example – Color segmentation

10 clusters
Practical Example – Color segmentation

15 clusters
Practical Example – Color segmentation

20 clusters
Practical Example – Color segmentation

Input Image
EM Algorithm for GMMs
• Idea:
• Objective function:
• Split optimization of the objective into to parts
• Algorithm:
• Initialize model parameters (randomly):
• Iterate until convergence:
• E-step
• Assign cluster probabilities (“soft labels”) to each sample
• M-step
• Find optimal parameters given the soft labels
Generalized EM
• Idea:
• Objective function:
• Split optimization of the objective into to parts
• Algorithm:
• Initialize model parameters (randomly):
• Iterate until convergence:
• E-step
• Assign cluster probabilities (“soft labels”) to each sample
• M-step
• Find optimal parameters given the soft labels
Generalized EM
• Idea:
• Objective function:
• Split optimization of the objective into to parts
• Algorithm:
• Initialize model parameters (randomly):
• Iterate until convergence:
• E-step
• Assign cluster probabilities (“soft labels”) to each sample
• M-step
• Find optimal parameters given the soft labels
Generalized EM
• Idea:
• Objective function:
• Split optimization of the objective into to parts
• Algorithm:
• Initialize model parameters (randomly):
• Iterate until convergence:
• E-step
• Assign cluster probabilities (“soft labels”) to each sample
• M-step
• Find optimal parameters given the soft labels
Generalized EM
• Idea:
• Objective function:
• Split optimization of the objective into to parts
• Algorithm:
• Initialize model parameters (randomly):
• Iterate until convergence:
• E-step
• Assign cluster probabilities (“soft labels”) to each sample
• M-step
• Find optimal parameters given the soft labels
Generalized M-step
• What is the objective function?

• GMM:

• General:
Exercise
• Consider a mixture of K multivariate Bernoulli distributions with
parameters , where

• Multivariate Bernoulli distribution:

• Question 1: Write down the equation for the E-step update

hint GMM: Answer:


Exercise
• Consider a mixture of K multivariate Bernoulli distributions with
parameters , where

• Multivariate Bernoulli distribution:

• Question 1: Write down the equation for the E-step update

hint GMM: Answer:


Exercise
• Consider a mixture of K multivariate Bernoulli distributions with
parameters , where

• Multivariate Bernoulli distribution:

• Question 2: Write down the EM objective:


Exercise
• Consider a mixture of K multivariate Bernoulli distributions with
parameters , where

• Multivariate Bernoulli distribution:

• Question 2: Write down the EM objective:


Exercise
• Multivariate Bernoulli distribution:

• EM objective:
Exercise
• Multivariate Bernoulli distribution:

• EM objective:
Exercise
• Multivariate Bernoulli distribution:

• EM objective:
Exercise
• Multivariate Bernoulli distribution:

• EM objective:
Exercise
• Question 3: Write down the M-step update

• Differentiate wrt.:
Exercise
• Question 3: Write down the M-step update

• Differentiate wrt.:
Exercise
• Question 3: Write down the M-step update

• Differentiate wrt.:
Summary
• EM algorithm is useful for fitting GMMs (or other mixtures) in an
unsupervised setting

• Can be used for:


• Clustering
• Classification
• Distribution estimation
• Outlier detection
Other unsupervised clustering techniques

Source: https://scikit-learn.org/stable/modules/clustering.html
Alternative for density estimation
• Kernel density estimation

Source: https://scikit-learn.org/stable/auto_examples/neighbors/plot_species_kde.html
References
• Lecture slides/videos
• https://www.python-
course.eu/expectation_maximization_and_gaussian_mixture_models.
php
• https://scikit-learn.org/stable/modules/clustering.html

You might also like