QUESTION=module, COUNT=8

Mixed ML Bag
QUESTION= singlecorrect, MARKS=1
True or False? A training dataset contains 100 instances of news articles that are labeled as
fake news and 500 instances that are labeled as real news. 80% accuracy with a trained
classifier on this dataset is a respectable accuracy.
OPTIONS=
1. True
2. False
ANSWER=B
QUESTION= singlecorrect, MARKS=1
Consider a point that is far away from the decision boundary of a logistic regression classifier
and is correctly classified. What will happen to the decision boundary if this point is removed
from the training set and the logistic regression classifier is retrained?
OPTIONS=
A. Decision boundary stays the same
B. Decision boundary changes
C. Cannot say
ANSWER=B
QUESTION=singlecorrect, MARKS=1
Let w* be the solution you obtain by optimizing the standard unregularized least-squares
objective for linear regression. What solution would you obtain if each of the input features (not
the labels) are scaled by a factor of 25 before estimating the regression weights?
OPTIONS=
A. w*
B. 25w*
C. w*/5
D. 5w*
E. w*/25
ANSWER=E
QUESTION=setcorrect, MARKS=1
Which of the following statements are true about Lasso least squares regression?
OPTIONS=
A. Weights are regularized with an L1 norm penalty
B. Weights have a Gaussian prior
C. Weights are typically sparse
D. Weights are regularized with an L2 norm penalty
ANSWER=A,C
QUESTION=numeric, MARKS=1
Consider a Naive Bayes classifier for a binary classification task on a training dataset containing
two attributes and . Let the two classes which are equally likely be
denoted by y1 and y2. For attribute A, you are given that P(A = 0 | y1) = 0.4, P(A = 1 | y1) = 0.2,
P(A = 0 | y2) = 0.6 and P(A = 1 | y2) = 0.3. For attribute B, you are given that P(B = 0 | y1) = 0.4,
P(B = 1 | y1) = 0.3, P(B = 0 | y2) = 0.1 and P(B = 1 | y2) = 0.3. According to the Naive Bayes model,
what is the posterior probability P(y2 | A = 0, B = 2)? Write your answer as a decimal rounded up
to two decimal points.
ANSWER=[0.75,0.75]
QUESTION= singlecorrect, MARKS=1
True or False? Consider two binary logistic regression classifiers C and D that are trained using
gradient descent and a very small learning rate. Despite C and D starting with different
initialization points, they are likely to converge to the same solution.
OPTIONS=
A. True
B. False
ANSWER=A
QUESTION= singlecorrect, MARKS=1
True or False? In comparison to linear regression without any regularization, ridge regression
usually has lower bias.
OPTIONS=
A. True
B. False
ANSWER=B
QUESTION= singlecorrect, MARKS=1
True or False? Changing the regularization coefficient in ridge regression can affect the
bias-variance tradeoff.
OPTIONS=
A. True
B. False
ANSWER=A

========================================================================
========

QUESTION=module, COUNT=3, CATEGORY=medium

Here, c is a parameter. The conditional probability density P(Y | X, c) can be written as:

Say we have a training set of size n, (Xi, Yi), i = 1,…,n and σ is known. Assume a prior on c such
that . Let c* and c” denote the MLE and MAP estimates of c, respectively.
QUESTION=singlecorrect, MARKS=1
As , what can we say about | c* - c” | ?
OPTIONS=
A. Increases
B. Decreases
C. Stays the same
ANSWER=B
QUESTION=singlecorrect, MARKS=1
As the number of training examples , what can we say about | c* - c” | ?
OPTIONS=
A. Increases
B. Decreases
C. Stays the same
ANSWER=B
QUESTION=singlecorrect, MARKS=2
What is the MAP estimate of c? Assume σ2=1.
OPTIONS=

A.

B.

C.

D.

E.
ANSWER=D

========================================================================
======

QUESTION=module, COUNT=2
Consider a logistic regression model P(y|x,w) for a binary classification task. In the training
dataset, x = (x1, x2), where x1, x2 are binary features and y is a deterministic function of the two
features: y = 1 if x1= x2 and y = 0 otherwise. With labels computed using this deterministic
function, say we have five training instances for each of the four combinations of (x1, x2), i.e. a
total of 20 training examples in the train set.
QUESTION=numeric, MARKS=1
What is the classification error of the training examples using the logistic regression classifier?
Write your answer as a decimal rounded up to two decimal points.
ANSWER=[0.25,0.25]
QUESTION=singlecorrect, MARKS=1
Which of the following features when used in combination with x1, x2 would give zero
classification error on the training examples?
OPTIONS=
A.

B.
C.
D.
ANSWER=C,D

========================================================================
========

QUESTION=paper, MARKS=3, CATEGORY=mle

Write down a formula for the maximum likelihood estimate (MLE) of a, given n i.i.d. samples
(x1,y1), …, (xn,yn), drawn from this distribution. Justify your answer.

========================================================================
========

QUESTION=module, COUNT=3
Basics of probability and linear algebra
QUESTION=singlecorrect, MARKS=1
Consider two nxn matrices A and B of rank r1 and r2, respectively. Let r be the rank of AB.
Choose the statement that is guaranteed to be correct.
OPTIONS=
A. r ≥ min(r1,r2)
B. r ≤ min(r1,r2)
C. r = r1+r2
D. None of the above
ANSWER=B
QUESTION=setcorrect, MARKS=1
Consider two nxn matrices A and B of rank r1 and r2, respectively. Let r be the rank of A+B.
Choose the statement that is guaranteed to be correct.
OPTIONS=
A. r ≥ max(r1,r2)
B. r ≤ max(r1,r2)
C. r = r1+r2
D. None of the above
ANSWER=D
QUESTION=numeric, MARKS=1
A random variable X is uniformly distributed over the interval [0,1]. What is the expectation
E[X2]? Write your answer as a decimal, rounded to two decimal places.
ANSWER=[0.33,0.33]