CS
5354/CS 4365 Intelligent Computing
Fall 2019, Test
2
Name:
____________________________
General comments:
- you are allowed up to 5 pages of
handwritten notes;
- place the last 4 digits of ID number on
each extra page;
- the main goal of most questions is to show
that you know the corresponding algorithms; so, if you are running
of time, just follow the few first steps of the corresponding
algorithm;
- each question will be graded on its own merit; so,
for example, if when answering to the first part of the question,
you got the wrong formula, but on the second part, you correctly
used this formula, you will get full credit for the second part.
Good luck!
1. Use Lagrange multiplier method to solve the following constraint
optimization problem: find a point on an ellipse
x12 + 2x22 = 1 which is
the furthest away from the point (4, 3).
2. Perform the first step of gradient descent to find the minimum
of a function 2(x1 + 2)2 +
3(x2 − 2)2. Start with a point
(0, 0) and use α = 0.1.
3. To find out which questions to answer, divide the last 2 digits
of your UTEP ID by 3.
- If the remainder is 0, answer questions 3, 6, 9, and 12.
- If the remainder is 1, answer questions 1, 4, 7, and 10.
- If the remainder is 2, answer questions 2, 5, 8, and 11.
3.1 Prove that the smoothest line connecting two points is a
straight line
3.2 What is machine learning: what do we know, and what do we want
3.3 What is the main objective of the traditional neural network?
3.4 Why in the traditional neural network, we only use linear and
1-input nonlinear neurons
3.5 Why cannot we have 1- or 2-layer neural networks?
3.6 Why we
use L-NL-L networks and not NL-L-NL ones?
3.7 Explain why 3-layer
neural networks are universal approximators
3.8 Which activation
function is used in traditional neural
networks? Why (biological answer).
3.9 What do we want to minimize when we train a neural network
3.10
What method do we use to minimize the corresponding objective
function?
3.11 Explain why gradient method is used for optimization
3.12 Explain how gradient descent leads to backpropagation formulas
for W0 and Wk.
Feel free to answer more questions, for extra credit.
4. (For extra credit) Provide a mathematical explanation of why
sigmoid is used in traditional neural networks.