Homework #1 solution key

Prayag
Neural networks and learning systems-I
February 25, 2019

Problem 1.
Solution. The output y is given by
y(n) = φ (wO1 y1 (n) + wO2 y2 (n)) (1)
where
 00 0

y1 (n) = φ x1 w11 + x2 w12 + y1 (n − 1)w11 + y2 (n)w12 (2)
 00 0

y2 (n) = φ x1 w21 + x2 w22 + y2 (n − 1)w22 + y1 (n)w21 (3)


Problem 2.

B B
z z

(a) (b)

Figure 1: (a) overlapping of convex hulls A and B. (b) linearly separable convex hulls

Solution. (a) From Figure 1(a) we see that the two convex hulls are intersecting. Assume
that there exist vectors w and b such that wT xn +b > 0 for all xn ∈ A and wT z n +b < 0
for all z n ∈ B. We know that A ∩ B 6= φ implies there exist at least one element
Nx
P Nz
P
y ∈ A ∩ B. We can write y = αn xn and y = βm z m for some α and β satisfying
n=1 m=1
the conditions as mentioned in the question. We look at wT y which is given by

Nx
αn wT xn + b > 0, for all xn ∈ A
 P


n=1
y = Nx (4)
αn wT z n + b < 0, for all z n ∈ B

 P

n=1

1
 From equation (4), we see that it is contradicting the assumption of existence of vectors
w and b.

(b) From Figure 1(b) we see that ∃ vectors wT , b such that wT xn + b > 0 for all xn ∈ A
and wT z n + b < 0 for all z n ∈ B. Now, we have to show A ∩ B = φ. We know that
Nx
P Nz
P
existence of w and b guarantees x = αn xn ∈
/ B and z = βm z m ∈
/ A implying
n=1 m=1
A ∩ B = φ.


Problem 3.

Solution. (a) Let x be the input with true class Ck . The total expected loss with Lkk = 0
for k = 1, . . . , N is given by
N X
X N Z
E(L) = Lkj P (x ∈ Ck | x) P (x)dx. (5)
k=1 j=1 C
j

Equation (5)

(b) The optimal rule for assigning class labels

Problem 4.

Solution. (a) The figure 2 shows the decision boundary for the AND and OR operations.

(a) (b)

Figure 2

(b) We know that the output of the XOR operation forms a non-linearly separable class
as shown in Figure 3. Therefore, it is not possible to separate the two classes using
perceptron.


2
 Figure 3: The two classes are non-linearly separable.

Figure 4: Full moon and crescent moon shapes with Rm = 5, Rin = 10, Rout = 15, and
D = 0.

Problem 5.

Solution. (a) The required data points are generated as shown in Figure 4

(b) The value of D is varied from 0 to 7 and the resulting perceptron decision boundary
is shown in Figures 5a-5h.

(c) With D = 7, five different initial weight vectors were chosen and the resulting per-
ceptron decision boundary is shown in Figures 6a-6e. The decision boundaries were
different with different initial conditions.

(d) With D = 7, the sequence in which the input is presented is randomized and the
resulting perceptron decision boundary is shown in Figures 7a-7e. We do not observe
any significant difference in the decision boundaries.

(e) A Gaussian noise with mean 0 and standard deviation ranging from 1 to 3.1 is added
to the dataset generated as shown in Figure 4 and the resulting perceptron decision
boundary is shown in Figures 8a-8d. The stopping criterion is error threshold set to
10−6 . The linear classification is not possible with the increase in standard deviation.

(f) The initial learning rate is varied from 0.1 to 1 in steps of 0.1. The initial weight vector
was set to [0, 0]T . The error threshold as stopping criterion is set to 10−6 . The number

3
 (a) (b) (c)

(d) (e) (f)

(g) (h)

Figure 5

of epochs required for convergence is 3 for all initial learning rates. This indicates that
the convergence is independent of the learning rate.


4
(a) (b) (c)

(d) (e)

Figure 6

(a) (b) (c)

(d) (e)

Figure 7

(a) (b) (c)

(d)

Figure 8