CS 771A: Intro To Machine Learning, IIT Kanpur (1 Feb 2022) Name Roll No Dept
CS 771A: Intro To Machine Learning, IIT Kanpur (1 Feb 2022) Name Roll No Dept
Name 20 marks
Roll No Dept. Page 1 of 2
Instructions:
1. This question paper contains 1 page (2 sides of paper). Please verify.
2. Write your name, roll number, department above in block letters neatly with ink.
3. Write your final answers neatly with a blue/black pen. Pencil marks may get smudged.
4. Don’t overwrite/scratch answers especially in MCQ – such cases will get straight 0 marks.
5. Do not rush to fill in answers. You have enough time to solve this quiz.
Q1. Write T or F for True/False (write only in the box on the right-hand side) (6 x 1 = 6 marks)
If 𝒫, 𝒬 ⊂ ℝ2 are two non-empty convex sets, then their union i.e., 𝒫 ∪ 𝒬 can never
1
be convex.
𝑓: ℝ → ℝ is convex, non-differentiable and −1 ∈ 𝜕𝑓(𝑥0 ) (i.e., -1 is a subgradient),
2
then 𝑓 must be decreasing around 𝑥0 i.e., for some 𝜖 > 0, 𝑓(𝑥0 − 𝜖) > 𝑓(𝑥0 + 𝜖).
If 𝒞 ≝ {(𝑥, 𝑦) ∈ ℝ2 : 2 ≤ 𝑥 + 𝑦 ≤ 3} is the set of all 2D vectors whose sum of
3
coordinates is between 2 and 3, then 𝒞 is a convex set.
If 𝐮, 𝐯 ∈ ℝ3 are two vectors such that ‖𝐮‖2 ≤ 1 and ‖𝐯‖2 < 2 then it is possible
4
that 𝐮⊤ 𝐯 > 2.5 for certain values of 𝐮, 𝐯.
If 𝐮, 𝐯 ∈ ℝ3 are two vectors such that ‖𝐮‖2 ≤ 1 and ‖𝐯‖2 < 2 then it is possible
5
that ‖𝐮 − 𝐯‖2 > 2.5 for certain values of 𝐮, 𝐯.
Classification models such as SVMs are more confident about their predictions on
6
data points that lie very close to their decision boundary.
Q2. (Sliding parabolas) Consider the two functions 𝑓(𝑥) = (𝑥 − 𝑎)2 + 𝑏 and 𝑔(𝑥) = 𝑥 4 + 𝑥 2 .
Find values of 𝑎, 𝑏 ∈ ℝ such that the
functions 𝑓, 𝑔 share a tangent at 𝑥 = 1. 𝑎= 𝑏=
For the values you find above, find out the
value of the function 𝑓 at 𝑥 = 1. 𝑓(1) =
Write down the equation of the (shared)
tangent of the functions at 𝑥 = 1. 𝑓(1) = 𝑔(1) =
Find values of 𝑎, 𝑏 ∈ ℝ such that the
functions 𝑓, 𝑔 share a stationary point 𝑥0
such that 𝑓(𝑥0 ) = 𝑔(𝑥0 ).
𝑎= 𝑏=
Note that your equation for the tangent must not contain variables like 𝑎, 𝑏 and should contain
only absolute constants. Write your answers only in the space provided. (2 + 1 + 1 + 2 = 6 marks)
Q3 (Robust regression) Training with the absolute loss can make regression models less sensitive
to outliers. Let us analyse a simplified version of the objective function used in this technique.
a. Consider the function 𝑓(𝑥) = |𝑏 ⋅ 𝑥 − 𝑎| where 𝑎, 𝑏, 𝑥 ∈ ℝ. For any value of 𝑎, 𝑏, 𝑥0 ∈ ℝ,
write down an expression for the (entire) subdifferential of 𝑓 at 𝑥0 . No derivation needed.
Page 2 of 2
1
b. Consider the function 𝑔(𝐰) = ‖𝐰‖22 + |𝐰 ⊤ 𝐱 − 𝑦| where 𝐰, 𝐱 ∈ ℝ𝑑 and 𝑦 ∈ ℝ. Write
2
down an expression for the (entire) subdifferential of 𝑔 at 𝐰 0 for any value of 𝐰 0 , 𝐱 ∈ ℝ𝑑
and 𝑦 ∈ ℝ. No derivation needed. Write your answer in the space given. (2 + 2 = 4 marks)
Your answer to part a.
Q4. (Vector line-up) Give examples of 4D vectors (fill-in the 4 boxes) with the following properties.
Any example will get full marks so long as it satisfies the property mentioned in the question part.
Your answers to the parts a, b, c, d may be same/different. (4 x 1 = 4 marks)
a. A vector 𝐯 ∈ ℝ4 such that 𝐯 ≠ 𝟎 and 𝐯 ⊤ 𝐮 = 0 where
𝐮 = (1, −1,1, −1) ∈ ℝ4 .