Scribd
Upload
14 views5 pages

IML19 Term1

The document is an examination paper for an Introduction to Machine Learning course, consisting of three main questions with various subparts. It covers topics such as classification types, decision tree algorithms, neural networks, overfitting, K-means clustering, Gaussian Mixture Models, and hyperparameters. Students are required to answer two questions within a specified time frame and demonstrate their understanding of machine learning concepts through calculations and explanations.

Uploaded by

Alexander Arzt
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
14 views5 pages

IML19 Term1

The document is an examination paper for an Introduction to Machine Learning course, consisting of three main questions with various subparts. It covers topics such as classification types, decision tree algorithms, neural networks, overfitting, K-means clustering, Gaussian Mixture Models, and hyperparameters. Students are required to answer two questions within a specified time frame and demonstrate their understanding of machine learning concepts through calculations and explanations.

Uploaded by

Alexander Arzt
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

PAPER C395

INTRODUCTION TO MACHINE LEARNING

Tuesday 17 March 2020, 10:00


Duration: 90 minutes
Post-processing time: 30 minutes
Answer TWO questions

Paper contains 3 questions


1a You are given three different problems below. For each problem, classify it as
either supervised learning, unsupervised learning, or reinforcement learning.
Also justify (in one sentence) why you classified each problem as such.

i) A book distributor has a collection of books, which it has classified into


different categories, e.g. “Young adults”, “Science fiction”, “Biography”,
and “Horror”. It wants to use this information to build a system to classify
its new products automatically.

ii) A supermarket has a database of its customers, and wants to automatically


discover and group its customers into different market segments to target
them separately.

iii) A group of aviation companies wants to develop a machine learning


algorithm which can predict the ( x, y) coordinates pinpointing the location
of a plane that has crashed. To develop this, the companies have collected
historical plane crash data, which include the coordinates of the planes’
crash sites.

b You are given the dataset below. Each row is a sample email annotated as spam
(or not), given whether or not a word appears in the email (0 indicates that the
word does not appear in the email, 1 indicates that it does).

# cash win debt home Spam?


1 0 1 0 0 no
2 1 1 1 0 yes
3 1 0 0 0 no
4 0 0 1 0 no
5 1 1 0 1 yes
6 1 1 1 1 yes
7 0 1 1 0 yes
8 1 0 1 1 no
9 0 0 0 0 no
10 1 1 0 0 yes

i) Using the Information Gain metric, which attribute will be selected as the
root node of a decision tree classifier? Please show all calculations
(including the Information Gain of all candidate nodes) to justify your
answer.

ii) Give one reason why one would need to prune a decision tree. Also
describe (in one or two sentences) how a validation set can be useful in
performing pruning.

c Imperial College London 2019 - 2020 Paper C395 Page 1 of 4


c Consider the training dataset below for a single-variable regression problem.

(i )
i x (i ) y (i ) d ( x ( q ) , x (i ) ) wq
1 1.5 3.16 ??? ???
2 2.3 1.45 ??? ???
3 3.0 1.07 ??? ???
4 3.8 2.01 ??? ???
5 4.9 4.51 ??? ???

i) At test time, you are given a query x (q) = 4.2.


(i )1
Given d( x (q) , x (i) ) = | x (q) − x (i) | and wq = , where | x |
d( x (q) ,x (i) )
indicates the absolute value of x, please complete the table above.

ii) Predict the output y(q) for x (q) = 4.2 using the k-nearest neighbours
regression algorithm with d( x (q) , x (i) ) as its distance measure, and
assuming k = 3. Show your calculation.

iii) Now predict the output y(q) for x (q) = 4.2 using the locally weighted
k-nearest neighbours regression algorithm. Use k = 3, the distance
(i )
measure d( x (q) , x (i) ), and the weights wq . Show your calculation.

The three parts carry, respectively, 30%, 40%, and 30% of the marks.

c Imperial College London 2019 - 2020 Paper C395 Page 2 of 4


2a Consider a fully-connected feedforward neural network for regression with 1
hidden layer and a single output neuron. Both the hidden and the output layers
use sigmoid activation. Mean squared error is used as the loss function for
optimisation.
Write out the necessary calculations for updating both the connection weights
and bias weights in the last layer using gradient descent. You are given the
matrix of input features X with N datapoints, output values Ŷ from the network,
the desired targets (labels) Y, and the current network weights.

b Explain the concept of overfitting. Name 3 methods you can use to deal with
overfitting and explain how each of them helps.

c A bank has developed a machine learning model for automatically identifying


fraudulent card transactions, which will then be manually reviewed. You run
some examples through the model and get the following results:

Example nr True label Predicted label


1 fraud real
2 real real
3 real fraud
4 real real
5 fraud real
6 real real
7 real real
8 fraud fraud
9 real real
10 real real
11 real real
12 fraud real
13 fraud fraud
14 real real
i) Construct the confusion matrix for this output.

ii) Calculate the classification accuracy, along with precision, recall and F1
for both classes.

iii) Analyse the results. Are there any issues? If so, which metrics identify
them?

The three parts carry, respectively, 40%, 30%, and 30% of the marks.

c Imperial College London 2019 - 2020 Paper C395 Page 3 of 4


3a Suppose at an update step, the K-means algorithm computes 3 cluster centroids:
µ1 = h−3, −1i, µ2 = h1, 2i, and µ3 = h−4, 1i. It then executes a cluster
assignment step. Assume that the algorithm uses a Euclidean distance measure.
To which cluster will the training example x (i) = h−2, 0i be assigned after the
cluster assignment step? Justify your answer by showing your calculations.

b Consider below the parameters θ = {πk , µk , σk2 : k = 1, 2, 3} of a univariate


Gaussian Mixture Model with 3 components that has been fitted to a set of
training examples, where πk is the mixing proportion of component k, µk the
mean of component k, and σk2 the variance for component k:

k πk µk σk2
1 0.5 -2 1
2 0.3 1 4
3 0.2 4 0.25

Suppose you are given an example x (i) = 0 at test time. Compute the probability
density for p( x (i) |θ ) given the parameters of the Gaussian Mixture Model above.
Show your calculations.
( x − µ )2
−
Hint: the Gaussian distribution is defined as: N ( x |µ, σ2 ) = √ 1 exp 2σ2
2πσ2

c When developing a neural network, which activation function and loss function
would you use in the output layer for the following applications? Justify your
decisions.

i) Predicting the temperature for tomorrow, based on the weather today.

ii) Generating text by predicting the next word in the sequence based on the
previous words.

iii) Detecting whether the camera image from a self-driving car contains a stop
sign.

d Hyperparameters are something that we need to deal with when designing


machine learning models.

i) Explain what hyperparameters are and give 2 examples.

ii) Given a dataset of 10,000 datapoints, how would you use it to find good
hyperparameters?

The four parts carry, respectively, 20%, 30%, 30%, and 20% of the marks.

c Imperial College London 2019 - 2020 Paper C395 Page 4 of 4

You might also like