Description
Question 1 (50%)
Using the K-Means clustering algorithm with minimum Euclidean-distance-based assignments
of samples to cluster centroids, segment the two attached color images into K ∈ {2,3,4,5} segments. As the feature vector for each pixel use a 5-dimensional feature vector consisting of normalized vertical and horizontal coordinates of the pixel relative to the top-left corner of the image,
as well as normalized red, green, and blue values of the image color at that pixel. Normalize
each feature by linearly shifting and scaling the values to the interval [0,1], such that the set of
5-dimensional normalized feature vectors representing each pixel are in the unit-hypercube [0,1]
5
.
For each K ∈ {2,3,4,5}, let the algorithm assign labels to each pixel; specifically, label lrc ∈
{1,…,K} to the pixel located at row r and column c. Present your clustering results in the form of
an image of these label values. Make sure you improve this segmentation outcome visualization
by using a contrast enhancement method; for instance, assign a unique color value to each label
and make your label image colored, or assign visually distinct grayscale value levels to each label
value to make best use of the range of gray values at your disposal for visualization.
Repeat this segmentation exercise using GMM-based clustering. For each specific K, use the
EM algorithm to fit a GMM with K components, and then use that GMM to do MAP-classification
style cluster label assignments to pixels. Display results similarly for this alternative clustering
method. Briefly comment on the reasons of any differences, if any.
Question 2 (50%)
In this exercise, you will train two support vector machine (SVM) classifiers and assess/compare
their test performances. These SVMs wil respectively have linear and spherically-symmetric Gaussian (shaped radial basis function) kernels. We will refer to them as Linear-SVM and GaussianSVM. The data vectors are two-dimensional real-valued. The data distributions for the two classes
are as follows: (1) data from class −1 are drawn from a Gaussian with zero-mean and identitycovariance-matrix; (2) data from class +1 are generated using a two-step procedure: a radius value
is drawn from a uniform distribution over the interval [2,3] and an angle value (in radians) is drawn
from a uniform distribution over the interval [−π,π]; these radius and angle values are converted
to Cartesian coordinates using the Polar-to-Cartesian coordinate transformation rule.
1
1. Generate a training set with 1000 independent samples from these two class distributions
with priors q− = 0.35 and q+ = 0.65; note that this does not mean 350 samples from one
class and 650 from the other – the class label needs to be randomly selected for each sample,
in accordance with this prior. Visualize your training data.
2. Using 10-fold cross-validation, and minimum probability of error as the objective, select the
hyper parameters for both Linear-SVM and Gaussian-SVM. For both classifiers, the constraint violation term weight (usually dwnoted by C; sometimes called the overlap penalty
weight; referred to as the box constraint parameter in Matlab’s fitcsvm) must be optimized.
For the Gaussian kernel, the scale parameter (usually denoted by σ, corresponds to the standard deviation, if this Gaussian was a probability distribution) needs to be optimized. Visualize your cross-validation process in search of optimal hyperparameter values. Report the
smallest probability of error estimate you get from cross-validation.
3. Using the best hyperparameters you identified, train your Linear-SVM and Gaussian-SVM
using all of the training dataset. Visualize classification results on training data, count the
erroneously classified samples and report the training dataset probability of error estimate.
4. Generate 1000 independent test samples from the same class distributions with the same
priors as in the training dataset. Apply the Linear-SVM and Gaussian-SVM classifiers to the
test data samples. Visualize the performance of your classifiers on the test dataset and report
your test probability of error estimate.
2