HWK 2
Wang: CSE 5526, Homework
1
CSE 5526 - Autumn 2013
Introduction to Neural Networks
Homework #2
Due Tuesday, Sept. 17
Grader: Kun Han
Office: 418 Caldwell, 688-3711
Office hours: 1:30-2:30 T & W
Email: hank@cse.ohio-state.edu (preferred)
Problem 1. (a) For the following training samples:
x
1
=(0, 0)
T
C
1
x
2
=(0, 1)
T
C
1
x
3
=(1, 0)
T
C
2
x
4
=(1, 1)
T
C
2
Plot them in input space. Apply the perceptron learning rule to the above samples one-at-
a-time to obtain weights that separate the training samples. Set to 0.5. Work in the
space with the bias as another input element. Use w(0) =(0, 0, 0)
T
. Write the expression
for the resulting decision boundary.
(b) XOR. For x
2
, x
3
C
1
and x
1
, x
4
C
2
, describe your observation when you
apply the perceptron learning rule following the same procedure as in (a).
Problem 2. The following figure shows the decision regions of four classes. Design a
classifier for these linearly inseparable classes, using a network of M-P neurons with
three output units. For class i (1 i 3), classification requires that y
i
=1, while y
j
=-1
for j i; Class 4 is recognized when y
i
=-1 for 1 i 3. (HINT: try a two-layer
feedforward network.)
Class 1
Class 2
Class 3
Class 4
0 2 5 8 - 4
2
4
5
2
1
x
x
Wang: CSE 5526, Homework
2
Problem 3. Given the following input points and corresponding desired outputs:
X ={-0.5, -0.2, -0.1, 0.3, 0.4, 0.5, 0.7}
D ={-1, 1, 2, 3.2, 3.5, 5, 6}
write down the cost function with respect to w (setting the bias to zero). Compute the
gradient at the point w =2 using both direct differentiation and LMS approximation
(average for all data samples in both cases), and see if they agree.
