Practice Session

Midsem Exam
Jun 15th, 2024 (Sat), 6PM, L19, L20
Only for registered students (regular + audit)
Assigned seating – to be announced on Piazza
Open notes (handwritten only)
No mobile phones, tablets etc
Bring your institute ID card
If you don’t bring it, you may have to spend precious
time during the exam getting verified separately
Syllabus:
All videos, slides, code linked on the course
discussion page (link below) till 10 Jun 2024 (Mon)
https://www.cse.iitk.ac.in/users/purushot/courses/ml/
2023-24-s/discussion.html
See GitHub for practice questions
True or False
For any vectors , we always have .
counts the number of non-zero coordinates of the vector
Suppose coordinate is nonzero in – then the coordinate must be non-
zero in either or or both.
Suppose coordinate is zero in – then the coordinate must be zero in
either or or both. – False since coordinate values can cancel.
True or False
The optimum for is always
Parallel Classifier
Create a feature map for some so that for
any , takes value if is in the dark cross-
hatched region and if is in the light dotted
region (see fig). E.g., is labelled while the
points and are both labelled . The lines in
the figure are and . We don’t care what
values are taken on points lying on these
two lines (as these are the decision
boundaries). is the all-ones vector.
CM to the rescue
Consider the following problem where and are constants and . Design a coordinate
minimization algorithm (choose coordinates cyclically) to solve the primal.

s.t., with and

If constraint would become
If constraint would become
If , constraint vanishes for
s.t., -- note that either or or both
Optimal DT
Melbo has a multiclass problem with three classes . There are 16
datapoints in total, each with a 2D feature vector . can take value
or . The table below describes each data point. All points are at the
root of a decision tree. Melbo wishes to learn a decision stump
based on the entropy reduction principle to split this node into two
children.
Optimal DT
What is the entropy of the root node?
Proportions are Entropy would be
What is the entropy of the two child if the split is done using the x
feature (x=0 becomes left child, x=1 becomes right child)?
(8,4,4)  (4,2,2) , (4,2,2)  , . Entropy
What is the reduction in entropy (i.e., H_root-H_children) if the split is
done using the x feature as described above?
0
Optimal DT
What is the entropy of the two child nodes if the split is done using
the y feature (y=0 becomes left child, y=1 becomes right child)?
(8,4,4)  (0,4,4) , (8,0,0)   =
What is the reduction in entropy (i.e., H_root-H_children) if the split
is done using the y feature as described above?
1
Doubts
Could we use Fourier transform to generate models? I believe that
would be more accurate while generating models for stocks and
similar types.
Yes, in fact deep learning can be seen as using learnt transforms
I had doubt that b is all the thing that is not coefficient of x or there
is some other criteria also that b should fulfill?
where and
When we talk about exact Barrier function in Lagrangian for
example if g(x)<=0 then we say r(x)=max alpha*g(x) (alpha is
greater than equal to zero) then this r(x) is value of alpha or it is
value of alpha*g(x)?