Artificial Neural Networks are computational models that are inspired by researches in neuroscience and biological cognitive functions. It provides an alternative computational paradigm to that first introduced by Von Neuman and has been successfully exploited in a wide variety of scientific and engineering disciplines, ranging from classification to function approximation. Development of a neural network model requires a thorough understanding of the basic principles and the underlying theory for their successful exploitation.
This course will give participants an introduction to the basic techniques of developing neural network models for a range of applications. Stronger emphasis is placed on the practical component, which are amply supported during the lectures. The participants are also exposed to a collection of data pre-processing techniques.
This course is aimed at scientists and engineers who wish to extend their problem-solving skills. No prior knowledge of the area is assumed, but participants should normally have a degree in a scientific discipline. Most of the practicals are based on MATLAB on UNIX workstations, and so familiarity of this environment is desirable.
The course includes:
Essentials of Neural Networks:
Historical background,
biological model of neuron, description of McCulloch-Pitts model, terminology,
flavour of range of applications, perceptron model.
Supervised Learning:
Single layer networks,
linear separability, perceptorn training algorithm, pocket algorithm adaptive
linear elements, least mean square learning rule and its
generalisation.
Multilayer Networks:
Hidden layers, backpropagation
algorithm, accelerating the learning process, quickprop algorithm. Adaptive
multilayer networks and network pruning algorithms. Radial Basis Function
Networks, decision regions, validation and novelty detection in MLPs, knowledge
discovery from MLPs.
Unsupervised Learning:
Simple competitive
learning, winner-take-all networks (Harming networks, Maxnet), learning vector
quantifiers, counter propagation networks, self-organising maps (Kohonen
networks).
Associative Models:
Characteristics of
Hopfield networks, discreet and continuous Hopfield networks, energy of the
networks, Boltzman machines, Hetero-associators.
Statistical Techniques:
Input data preparation, discriminant
analysis - linear, quadratic and logistic; cross validation, principal
component analysis.
Practicals:
Software written in
MATLAB (with Neural Net Toolbox) in used predominantly during the tutorials.
Additional software includes Neural Ware Professional and software written in
C.
The course lectures will be given by the teaching and research staff of the Applied Mathematics & Operational Research Group under the direction of Dr Venkat V S S Sastry with the assistance of other colleagues. External Speakers may give lectures on advanced topics.