CSCI 567 Homework #4 Programming Assignment solution


Original Work ?


5/5 - (1 vote)

Problem 1 K-means Clustering

Recall that for a dataset x1, . . . , xN ∈ RD, the K-means distortion objective is
J({µk}, {rik}) = 1


rikkµk − xik
where µ1, . . . , µK are centroids of the K clusters and rik ∈ {0, 1} represents whether example i belongs to
cluster k.

Clearly, fixing the centroids and minimizing J over the assignment give
rˆik =
1 k = arg mink
0 kµk
0 − xik
0 Otherwise.

On the other hand, fixing the assignment and minimizing J over the centroids give
µˆ k =



What the K-means algorithm does is simply to alternate between these two steps. See Algorithm 1 for
the pseudocode.

Algorithm 1 K-means clustering algorithm
1: Inputs:
An array of size N × D denoting the training set, x
Maximum number of iterations, max iter
Number of clusters, K
Error tolerance, e

2: Outputs:
Array of size K × D of means, {µk}
Membership vector R of size N, where R[i] ∈ [K] is the index of the cluster that
example i belongs to.

3: Initialize:
Set means {µk}
to be K points selected from x uniformly at random (with
replacement), and J to be a large number (e.g. 1010)
4: repeat
5: Compute membership rik using eq. 2
6: Compute distortion measure Jnew using eq. 1
7: if |J − Jnew| ≤ e then
8: end if
9: Set J := Jnew
10: Compute means µk using eq. 3
11: until maximum iteration is reached

1.1 Implementing K-means clustering algorithm Implement Algorithm 1 by filling out the TODO parts
in class KMeans of file Note the following:
• Use numpy.random.choice for the initialization step.
• If at some iteration, there exists a cluster k with no points assigned to it, then do not update the
centroid of this cluster for this round.

• While assigning a sample to a cluster, if there’s a tie (i.e. the sample is equidistant from two centroids),
you should choose the one with smaller index (like what numpy.argmin does).
After you complete the implementation, execute bash command to run k-means on toy

You should be able to see three images generated in plots folder. In particular, you can see
toy dataset predicted labels.png and toy dataset real labels.png and compare the clusters identified by the algorithm against the real clusters. Your implementation should be able to recover the
correct clusters sufficiently well. Representative images are shown in fig. 2. Red dots are cluster centroids.

Note that color coding of recovered clusters may not match that of correct clusters. This is due to mis-match
in ordering of retrieved clusters and correct clusters (which is fine).
(a) Predicted Clusters (b) Real Clusters
Figure 2: Clustering on toy dataset

1.2 Image compression with K-means In the next part, we will look at lossy image compression as an
application of clustering. The idea is simply to treat each pixel of an image as a point xi
, then perform
K-means algorithm to cluster these points, and finally replace each pixel with its centroid.
What you need to implement is to compress an image with K centroids given. Specifically, complete the
function transform image in the file

After your implementation, execute bash again and you should be able to see an image
baboon compressed.png in the plots folder. You can see that this image is distorted as compared to
the original baboon.tiff.

1.3 Classification with k-means Another application of clustering is to obtain a faster version of the nearest neighbor algorithm. Recall that nearest neighbor evaluates the distance of a test sample from every
training point to predict its class, which can be very slow. Instead, we can compress the entire training
set to just the K centroids, where each centroid is now labeled as the majority class of the corresponding
cluster. After this compression the prediction time of nearest neighbor is reduced from O(N) to just O(K)
(see Algorithm 2 for the pseudocode).

Algorithm 2 Classification with K-means clustering
1: Inputs:
Training Data : {X,Y}
Parameters for running K-means clustering

2: Training:

Run K-means clustering to find centroids and membership (reuse your code from Problem 1.1)
Label each centroid with majority voting from its members. i.e. arg maxc ∑i
rikI{yi = c}

3: Prediction:
Predict the same label as the nearest centroid (that is, 1-NN on centroids).
Note: 1) break ties in the same way as in previous problems; 2) if some centroid doesn’t contain any
point, set the label of this centroid as 0.

Complete the fit and predict function in KMeansClassifier in file Once completed,
run to evaluate the classifier on a test set. For comparison, the script will also print accuracy of
a logistic classifier and a nearest neighbor classifier. (Note: a naive K-means classifier may not do well but
it can be an effective unsupervised method in a classification pipeline [2].)

Problem 2 Gaussian Mixture Model

Next you will implement Gaussian Mixture Model (GMM) for clustering and also generate data after learning the model. Recall the key steps of EM algorithm for learning GMMs on Slide 52 of Lec 8 (we change the
notation ωk
to πk
γik =
πkN (xi
; µk
, Σk
∑k πkN (xi
; µk
, Σk
Nk =

γik (5)
µk =

i=1 γikxi
Σk =

i=1 γik(xi − µk
)(xi − µk
πk =

Algorithm 3 provides a more detailed pseudocode. Also recall the incomplete log-likelihood is

ln p(xn) =


πkN (xi
; µk
, Σk
) =



(xi − µk

(xi − µk


2.1 Implementing EM Implement EM algorithm (class Gaussian pdf, function fit and function com
pute log likelihood) in file to estimate mixture model parameters. Please note the following:
• For K-means initialization, the inputs of K-means are the same as those of EM’s.
• When computing the density of a Gaussian with covariance matrix Σ, use Σ
0 = Σ + 10−3
I when Σ is
not invertible (in case it’s still not invertible, keep adding 10−3
I until it is invertible).

After implementation execute bash command to estimate mixture parameters for toy dataset.
You should see a Gaussian fitted to each cluster in the data. A representative image is shown in fig. 3.
We evaluate both initialization methods and you should observe that initialization with K-means usually
converges faster.

Figure 3: Gaussian Mixture model on toy dataset
Algorithm 3 EM algorithm for estimating GMM parameters
1: Inputs:
An array of size N × D denoting the training set, x
Maximum number of iterations, max iter
Number of clusters, K
Error tolerance, e
Init method — K-means or random

2: Outputs:
Array of size K × D of means, {µk}
Variance matrix Σk of size K × D × D
A vector of size K denoting the mixture weights, pi k

3: Initialize:
• For “random” init method: initialize means uniformly at random from [0,1)
for each dimension (use numpy.random.rand), initialize variance to be
identity matrix for each component, initialize mixture weight to be uniform.
• For “K-means” init method: run K-means, initialize means as the centroids
found by K-means, and initialize variance and mixture weight according to
Eq. (7) and Eq. (8) where γik is the binary membership found by K-means.
4: Compute the log-likelihood l using Eq. (9)
5: repeat
6: E Step: Compute responsibilities using Eq. (4)
7: M Step:
Estimate means using Eq. (6)
Estimate variance using Eq. (7)
Estimate mixture weight using Eq. (8)
8: Compute new log-likelihood lnew
9: if |l − lnew| ≤ e then
10: end if
11: Set l := lnew
12: until maximum iteration is reached

2.2 Implementing sampling We also fit a GMM with K = 30 using the digits dataset. An advantage of
GMM compared to K-means is that we can sample from the learned distribution to generate new synthetic
examples which look similar to the actual data.

To do this, implement sample function in which uses self.means, self.variances and
self.pi k to generate digits. Recall that sampling from a GMM is a two step process: 1) first sample a
component k according to the mixture weight; 2) then sample from a Gaussian distribution with mean µk
and variance Σk
. Use numpy.random for these sampling steps.

After implementation, execute bash again. This should produce visualization of means µk
and some generated samples for the learned GMM. Representative images are shown in fig. 4.
(a) Means of GMM learnt on digits (b) Random digits sample generated from GMM
Figure 4: Results on digits dataset


[1] sklearn.datasets.digits
[2] Coates, A., & Ng, A. Y. (2012). Learning feature representations with k-means. In Neural networks:
Tricks of the trade (pp. 561-580). Springer, Berlin, Heidelberg.