Honours Adv. Topic: Artificial Neural Networks1st Semester 2001
Honours Advanced Topic: Artificial Neural Networks
Semester 1, 2001
Written Examination
INSTRUCTIONS
This an open book examination.
Time: 3 hours and 10 minutes (reading time).
There are 5 questions.
Answer all questions.
Question 1 (18 marks)
a)(8 marks)
Describe what you know about the structure of a biological neuron and learning in biological systems.
Discuss the McCulloch-Pitt model of a biological neuron.Is it a linear or non-linear system? Explain your answer. What about biological neurons?
b)(6 marks)
The most common application of neural networks is pattern recognition.
What are the main stages in the pattern recognition process? Describe with examples, the concepts of feature vectors and discriminant functions in this context.
c)(4 marks)
Describe with example, what is meant by the standard basis vectors of a multi-dimensional space.
What are the implications of two pattern vectors being orthogonal?
Give examples of two pairs of such patterns, one of them binary and the other continuous valued.
Question 2 (20 marks)
a)(6 marks)
What is an ADALINE? Describe the learning rule used for an ADALINE.
Describe with an example problem the major drawback of a single layer of perceptrons.
b)(8 marks)
Derive the formula used for adjusting weights during the training of a multilayer perceptron.
Discuss the statement - any function can be represented by a multilayer perceptron for no more than three layers.
c)(6 marks)
Describe the concept of the error surface in a neural network
Discuss the techniques for overcoming learning difficulties in multilayer perceptrons.
Question 3 (21 marks)
a)(7 marks)
Describe the architecture of the Kohonen network and the formation of the so-called topological map through training.
b)(8 marks)
Discuss the characteristics of the Hopfield model followed by an account of its training and recognition stages.
c)(6 marks)
Justify the setting of connection weights in the training stage of the Hopfield net by deriving an expression for the weight term w from the Hopfield net energy function.
Question 4 (21 marks)
a)(6 marks)
Discuss simulated annealing and its use in assisting convergence to global minima in the Boltzmann machine.
b)(8 marks)
What is known as the stability-plasticity problem? Explain how this problem is solved in the neural network model proposed by Grossberg and Carpenter.
Discuss the limitations of the ART paradigm in general.
c)(7 marks)
Describe with a suitable diagram what fuzzy cognitive maps are.
How can FCMs be used to simulate dynamic systems and make predictions on their future states?
Question 5 (20 marks)
a)(6 marks)
Describe the architecture and operation of the counterpropagation network with emphasis on the training process. Compare counterpropagation and backpropagation networks.
b)(4 marks)
How does problem solving using an artificial neural network differ from that using a digital computer?
What types of problems are not considered suitable for neural nets?
c)(4 marks)
Discuss the decisions to be made at the node and network levels, during the design phase of an artificial neural network.
d)(6 marks)
Discuss the application of artificial neural networks in business or industry.
END OF PAPER
1