Professional Documents
Culture Documents
Exam 2004
Exam 2004
Instructions
There are SIXTY questions (worth up to 60 marks). The exam mark (maximum 60) will be added to the
mark obtained in the laborations (maximum 5). The total pass mark for the course is 35 out of 65.
For each question, please select a maximum of ONE of the given answers (either A, B, C, D or E).
You should select the one answer that represents the BEST possible reply to the question (in some cases,
there may be no obvious “wrong” answers, but one answer should always be better than the others).
Every time you select the correct answer, you will be awarded +1 mark. However, every time you select
an incorrect answer, a penalty score will be subtracted from your total mark. This penalty depends on
the number of possible answers to the question, as follows:
2 +1 −1
3 +1 − 12
4 +1 − 13
5 +1 − 14
If you do not give any answer to a question, no marks will be added to your total score and there will be
no penalty. If you give more than one answer to a question, this will be counted as an incorrect answer.
So please be very careful, and make sure that ONLY one letter (A or B or C or D or E) is visible in each
of your written answers. Please write your answers very clearly so that they can be read by an average
examiner!
1
Artificial Neural Networks – Examination, March 2004 2
Questions
1 Pattern Recognition
What is a class in a pattern recognition problem?
A. An input variable.
B. A hidden variable or weight.
C. An output variable.
2 Pattern Recognition
Is the following statement true or false? “In unsupervised learning, there is a target output vector which
tells the pattern recognition system the correct class for a given input vector.”
A. TRUE.
B. FALSE.
B. A plane.
C. A curve.
D. A hyperplane.
E. A hypercurve.
Artificial Neural Networks – Examination, March 2004 3
12 Biological Neurons
What are the advantages of biological neural networks (BNNs) compared to conventional computers?
13 Biological Neurons
Which of the following statements is the best description of Hebb’s learning rule?
A. “If a particular input stimulus is always active when a neuron fires then its weight should be
increased.”
B. “If a stimulus acts repeatedly at the same time as a response then a connection will form between
the neurons involved. Later, the stimulus alone is sufficient to activate the response.”
C. “The connection strengths of the neurons involved are modified to reduce the error between the
desired and actual outputs of the system.”
Artificial Neural Networks – Examination, March 2004 5
18 Perceptrons
A simple perceptron has two input units, a unipolar step function, weights w1 = 0.2 and w2 = −0.5, and
a threshold θ = −0.2 (θ can therefore be considered as a weight for an extra input which is always set to
T
-1). What is the actual output of this perceptron for the input pattern x = [1, 1] ?
A. -1.
B. 0.
C. 1.
19 Perceptrons
The perceptron in question 18 is trained using the learning rule 4w = η (d − y) x, where x is the input
vector, η is the learning rate, w is the weight vector, d is the desired output, and y is the actual output.
What are the new values of the weights and threshold after one step of training with the input vector
T
x = [0, 1] and desired output 1, using a learning rate η = 0.2?
A. w1 = 0.0, w2 = −0.5, θ = 0.0.
B. w1 = 0.2, w2 = −0.3, θ = 0.0.
C. w1 = 0.2, w2 = −0.3, θ = −0.4.
D. w1 = 0.2, w2 = −0.5, θ = −0.2.
E. w1 = 0.4, w2 = −0.5, θ = −0.4.
20 Perceptrons
What is the equation of the decision line for the perceptron in question 18 (before training)?
A. x2 = 0
B. x2 = 0.667x1
C. x2 = 0.400x1 + 0.400
D. x2 = 0.667x1 + 1.333
E. x2 = 0.800x1 + 0.800
21 Perceptrons
What is the equation of the decision line for the perceptron in question 19 (after training)?
A. x2 = 0
B. x2 = 0.667x1
C. x2 = 0.400x1 + 0.400
D. x2 = 0.667x1 + 1.333
E. x2 = 0.800x1 + 0.800
Artificial Neural Networks – Examination, March 2004 7
22 Perceptrons
A single-layer perceptron has 3 input units and 3 output units. How many weights does this network
have?
A. 6
B. 9
C. 18
D. 25
23 Perceptrons
In the following network, all of the units have binary inputs (0 or 1), unipolar step functions and binary
outputs (0 or 1).
6
4
6@
I
@
@
@
w43
3
w41 w42
J
]J
J
w31 w32J
J
1 2
6 6
The weights for this network are w31 = 1, w32 = 1, w41 = 1, w42 = 1 and w43 = −2. The threshold of the
hidden unit (3) is 1.5 and the threshold of the output unit (4) is 0.5. The threshold of both input units
(1 and 2) is 0.5, so the output of these units is exactly the same as the input. Which of the following
Boolean functions can be computed by this network?
A. AND.
B. OR.
C. XOR.
D. All of the above answers.
E. None of the above answers.
24 Perceptrons
Is the following statement true or false? “A multi-layer perceptron with separate layers of weights could
perform exactly the same function as the neural network in question 23.”
A. TRUE.
B. FALSE.
Artificial Neural Networks – Examination, March 2004 8
25 ADALINE
Which of the following statements is NOT true for an ADALINE network?
A. There are separate phases for training and testing.
B. The weight vector is constantly adapted.
J
v21J H y1
w12
j
H
J *
H 1
H
J
JH
HH
v12 J
J
H w21
J
HHj
J
x2 - J I2 - z
2
H
HH J
*
v22 H
H
J
J w 22
HHJ^
J
j J
v13
H2 H
Jw31
y2 HH
*
v23 w32 H J
H
JJ
HHj ^
x3 -
I3 - z
3
T T
The input vector to the network is x = [x1 , x2 , x3 ] , the vector of hidden layer outputs is y = [y1 , y2 ] ,
T T
the vector of actual outputs is z = [z1 , z2 , z3 ] , and the vector of desired outputs is t = [t1 , t2 , t3 ] . The
network has the following weight vectors:
−2.0 1.0
1.0 0.5 0.3
v1 = 2.0 , v2 = 1.0 , w1 = , w2 = and w3 = .
−3.5 −1.2 0.6
−2.0 −1.0
4wkj = ηδk yj
where 4wkj is the change to the weights from unit j to unit k, η is the learning rate, δk is the error term
for unit k, and yj is the j th input to unit k. In the backpropagation algorithm, what is the error term δk ?
A. δk = f 0 (Hj )(tk − zk ).
B. δk = f 0 (Ik )
P
k δk wkj .
C. δk = f 0 (Ik )(tk − zk ).
D. δk = f 0 (Hj )
P
k δk wkj .
4vji = ηδj xi
where 4vji is the change to the weights from unit i to unit j, η is the learning rate, δj is the error term
for unit j, and xi is the ith input to unit j. In the backpropagation algorithm, what is the error term δj ?
A. δj = f 0 (Hj )(tk − zk ).
B. δj = f 0 (Ik )
P
k δk wkj .
C. δj = f 0 (Ik )(tk − zk ).
D. δj = f 0 (Hj )
P
k δk wkj .
J . ..
.
.. .
. ...
. .
−3.5
J . . .
J. . . . . ...
. . . ...
. ..
. ..
.
. . . . ..
.. . .
..
.
J . .
. . ..
.. . .
. ..
.. J . . ..
HH .
.
. J
... *
I2 - z = 0.36
2
. .
.
HH .
. −1.2
^
. J .
.
.
.
. H J
.
.
. HH j .
. .
.
. H2
.
.
.
.
*
HH
y 2
.
.
.
. H .
.
0.6 .
. H .
. H .
. .
. H .
. j
H .
.
I3 - z = 0.57
3
For the output units, the derivative of the sigmoid function can be rewritten as
What is the error for hidden unit 2 given that its activation for the pattern being processed is currently
y2 = 0.50?
A. δhidden 2 = 0.2388
B. δhidden 2 = 0.1475
C. δhidden 2 = 0.0534
D. δhidden 2 = −0.1233
E. δhidden 2 = −0.1672
(Assume exactly the same weights, activation functions and bias values as described in question 30, and
exactly the same output vectors t and z as described in the previous question.)
B. The trained network can be split into two machines: the first layer of weights compresses the input
pattern (decoder), and the second layer of weights reconstructs the full pattern (encoder).
C. After training, the hidden units give a representation that is equivalent to the principal components
of the training data, removing redundant parts of the input data.
D. The data compression can be useful for pre-processing the data for training another neural network.
E. It is important to use smooth, non-linear activation functions in the hidden units.
Artificial Neural Networks – Examination, March 2004 13
39 Unsupervised Learning
Which of the following statements is NOT true for hard competitive learning (HCL)?
A. There is no target output in HCL.
B. The neurons have unipolar step functions.
C. During testing, HCL is equivalent to a minimum distance classifier.
D. The input vectors are often normalized to have unit length — that is, k x k= 1.
E. During training, the weights of the winning unit k are adapted by 4wk = η (x − wk ), where x is
the input vector.
40 Unsupervised Learning
Which of the following statements is NOT true for a self-organising feature map (SOM)?
A. There are two phases during training: soft competitive learning and fine tuning.
B. The topological mapping in a well trained SOM preserves the probability distribution of the training
data.
C. The units are arranged in a regular geometric structure such as a 2-dimensional square or a 1-
dimensional ring.
D. The number of cluster units can be increased by adding new units if the distribution of the input
data changes over time.
E. The learning rate for the neighbours is a decreasing function of the distance from the winning unit
in the geometric structure.
41 Unsupervised Learning
An input vector x and two prototype vectors p1 and p2 are given by
T
x = [−1.40, 2.30, 0.20]
T
p1 = [−1.00, 2.20, 0.10]
T
p2 = [−4.00, 7.00, 0.60]
42 Unsupervised Learning
Adapt the weight vector of the winning prototype in question 41 according to the SOM learning algorithm
with a learning rate of 0.5. What is the new weight vector?
T
A. pwinner = [−0.80, 2.15, 0.05]
T
B. pwinner = [−1.20, 2.25, 0.15]
T
C. pwinner = [−2.70, 4.65, 0.40]
T
D. pwinner = [−5.30, 9.35, 0.80]
Artificial Neural Networks – Examination, March 2004 14
43 Unsupervised Learning
A SOM can be used for fault detection in an industrial process. The network is trained using only examples
of the “normal” state of the process. During testing, a new input pattern x is presented to the network.
Which method is used to decide whether there is a possible fault in the process?
A. There may be a fault if the squared Euclidean distance from the input pattern x to the best matching
unit in the SOM is high.
B. There may be a fault if the squared Euclidean distance from the input pattern x to the best matching
unit in the SOM is low.
44 Unsupervised Learning
Is the following statement true or false? “An ensemble of SOMs (ESOM) can be used as a classifier. One
SOM is trained for each class in the training data. During testing, a new input pattern x is presented to
all of the SOMs, and the SOM with the highest distance to its best matching unit will be the winner.”
A. TRUE.
B. FALSE.
45 Associative Memory
How many hidden layers are there in an autoassociative Hopfield network?
A. None (0).
B. One (1).
C. Two (2).
46 Associative Memory
A Hopfield network has 20 units. How many adjustable parameters does this network contain?
A. 95
B. 190
C. 200
D. 380
E. 400
Artificial Neural Networks – Examination, March 2004 15
47 Associative Memory
Calculate the weight matrix for a Hopfield network to store the pattern 1 −1 1 −1 .
0 −1 −1 −1
−1 0 1 1
A. W = −1 1
0 1
−1 1 1 0
0 1 −1 −1
1 0 −1 −1
B. W = −1 −1 0
1
−1 −1 1 0
0 1 1 −1
1 0 1 −1
C. W =
1 1 0 −1
−1 −1 −1 0
0 −1 1 −1
−1 0 −1 1
D. W = 1 −1 0 −1
−1 1 −1 0
48 Associative Memory
Calculate the weight matrix for a Hopfield network to store two patterns 1 −1 1 −1 and −1 −1 −1 1 .
0 2 0 −2
2 0 0 −2
A. W = 0
0 0 0
−2 −2 0 0
0 −2 0 −2
−2 0 0 2
B. W = 0
0 0 0
−2 2 0 0
0 0 2 −2
0 0 0 0
C. W = −2 0 0 2
−2 0 2 0
0 0 2 −2
0 0 0 0
D. W = 2 0 0 −2
−2 0 −2 0
49 Associative Memory
The trained Hopfield network in question 48 is used for recall. A pattern −1 −1 1 −1 is presented
to the network, then the nodes of the network are updated until a steady state is reached. What is the
final state of the network?
A. −1 −1 1 −1
B. 1 −1 1 −1
C. −1 −1 −1 1
D. −1 1 −1 1
Artificial Neural Networks – Examination, March 2004 16
50 Genetic Algorithms
Is the following statement true or false? “A genetic algorithm will always increase the overall fitness of
the population.”
A. TRUE.
B. FALSE.
51 Genetic Algorithms
Which genetic operator can introduce new genetic information into the population?
A. Crossover.
B. Elitism.
C. Mutation.
D. Reproduction.
E. Sharing.
52 Genetic Algorithms
Which genetic operator is used to decrease the fitness of individuals that are too similar to other members
of the population?
A. Crossover.
B. Elitism.
C. Mutation.
D. Reproduction.
E. Sharing.
53 Genetic Algorithms
Which genetic operator is used to combine the genetic information from two parents to produce two
offspring?
A. Crossover.
B. Elitism.
C. Mutation.
D. Reproduction.
E. Sharing.
54 Laborations
What does the following command in MATLAB do?
>> y = sim(net, x);
55 Applications
Is the following statement true or false? “Learning produces changes within an agent that over time
enables it to perform more effectively within its environment.”
A. TRUE.
B. FALSE.
56 Applications
Which type of learning system is used for adaptive noise cancellation in telephone circuits?
A. A linear feedforward network.
B. A multi-layer feedforward network.
C. A simple recurrent network.
D. A self-organising feature map.
E. A genetic algorithm.
57 Applications
Which type of learning system can be used for path planning in mobile robots?
A. A linear feedforward network.
B. A multi-layer feedforward network.
C. A simple recurrent network.
D. A self-organising feature map.
E. A genetic algorithm.
58 Applications
Which type of learning system was used for gesture recognition on a mobile robot?
A. A linear feedforward network.
B. A multi-layer feedforward network.
59 Applications
The following schematic diagram shows the neural network that was used to train a mobile robot with an
electronic nose to estimate the direction to an odour source.
60 Applications
What is the usual sequence of events for training the neural network in question 59?
Answers
1. C +1 − 12
2. B +1 −1
3. A +1 − 14
4. E +1 − 14
5. C +1 − 12
6. A +1 −1
7. A +1 − 12
8. C +1 − 12
9. A +1 − 12
10. A +1 − 12
11. B +1 − 12
12. B +1 − 13
13. A +1 − 12
14. A +1 −1
15. A +1 − 13
16. E +1 − 14
17. D +1 − 14
18. B +1 − 12
19. C +1 − 14
20. C +1 − 14
21. D +1 − 14
22. B +1 − 13
23. C +1 − 14
24. A +1 −1
25. A +1 − 14
26. E +1 − 14
27. B +1 −1
28. C +1 − 13
29. D +1 − 13
30. A +1 − 14
31. D +1 − 14
32. C +1 − 13
33. D +1 − 13
34. C +1 − 14
35. C +1 − 13
36. D +1 − 14
37. E +1 − 14
38. B +1 − 14
39. B +1 − 14
40. D +1 − 14
41. A +1 −1
42. B +1 − 13
43. A +1 −1
44. B +1 −1
45. A +1 − 12
46. B +1 − 14
47. D +1 − 13
48. D +1 − 13
49. B +1 − 13
50. B +1 −1
Artificial Neural Networks – Examination, March 2004 20
51. C +1 − 14
52. E +1 − 14
53. A +1 − 14
54. C +1 − 13
55. A +1 −1
56. A +1 − 14
57. E +1 − 14
58. B +1 − 14
59. D +1 − 13
60. B +1 − 14