Instructor: Dr Robert Kozma (rkozma@memphis.edu or kozma@socrates.berkeley.edu )
                  Dept of Mathematical Sciences / Computer Science and Institute of Intelligent Systems

Time: Tuesday and Thursday, 5:30 - 6:55 p.m.

Motivation

In the past few decades research into neural networks grew from a narrow field encompassing clever connectionist architectures and optimization methods into a powerful tool that addresses key issues of knowledge processing in artificial/ computational and biological systems. The founding fathers of digital computers in the '40s and '50s in the last century, starting with John von Neumann, emphasized the importance of learning from understanding brain behavior when developing computer designs. Neural networks provide a unified view and general methodology that can be successfully implemented in a wide variety of research/ practical fields. Nowadays, intelligent neural network systems find their applications in

human-computer interfaces, speech processing and natural language understanding,

sensory information processing and cognitive processes,

intelligent and adaptive control and robotics,

image processing and pattern recognition, time series analysis and forecasting

including financial applications, environmental and climate changes, biomedical fields, and others.

In the first part of the course, basic methods of neural networks and connectionist science are introduced, starting with the multilayer perceptron and its universal mapping property, radial basis function networks, Hopfield nets; various supervised and unsupervised learning and optimization methods, as backpropagation, Hebbian learning; recurrent network architectures and related learning and convergence properties, including backpropagation through time.

We consider neural networks as (usually big) graphs and outline the importance of methods of discrete mathematics, combinatorics, and graph theory in neural network studies. Static and dynamic memory principles will be introduced. Special emphasis is given to various structural learning algorithms that allow the use of neural networks for data mining and knowledge discovery.

In the second part of the course, various application fields will be dealt with that cover the field of interest of the students (examples of possible topics are given above). Students will work on a project of their own choice, in which they will implement a solution to a practical problem using neural network principles learned at the class.