MIN 601601: Application of Artificial Neural Networks

Instructor: Rajive Ganguli, Ph.D., P.E., C.O.I

Assistant Professor of Mining Engineering

Office: 317 Duckering Office Hours: Tue: 2 - 4 PM

Tel: 474-7631 email: ffrg@uaf.edu

Course : The course covers basic neural network architectures, learning rules,

Description training methods and practical applications. Training and application issues typical of earth sciences problems are discussed. Some topics require mathematical analysis. Genetic algorithms and use of network ensembles will be briefly presented. Boosting and radial basis functions will be presented if time permits. The outcome of this course is a basic understanding of the development and application of artificial neural networks for earth science problems.

Prerequisites: i) Graduate standing in engineering ii) Ability to program matrices into computer code, knowledge of MATLAB (or another matrix language) a plus. RECOMMENDED: MIN 408, MIN 635, MATH 202, MATH 314

Textbook: Neural Network Design (ISBN: 0-534-94332-2)

By Hagan, Demuth and Beale, PWS Publishing Company

Grading: Homework: 60%

- You will have four to six homework assignments. These assignments will be programming and numerical intensive

Project: 40%

- You will have to design and code a neural network for a problem of your choice. You will have to come up with your own problem and present it to me as an initial proposal. If the proposal is acceptable to me, you will then continue with the neural network design for the problem. You will be required to present the design in a report as well as present it in class.

Grading: A >=90, 90>B>=80, 80>C>=70, 70>D>=60, 60>F>=0

Policy: Late work will NOT be accepted. Announcements on course web page or classroom or via email is considered official. It is the student’s responsibility to stay informed.

Course Content (numbers in parenthesis indicate approximate duration in weeks):

- Introduction to neural networks and architecture (11)

o Basic structure, introduction to Perceptron, Hamming network and Hopfield network.

- Perceptron Learning Rule (0.5)

o Developing a learning rule for a perceptron (single neuron case and multiple neuron case)

- Supervised Hebbian Learning (0.5)

- Review Fundamentals of Optimization (1.0)

o Taylor series, first order and second order optimality, quadratic functions, optimization methods

- Widrow-Hoff Learning (0.5)

o ADALINE network, mean square errors, LMS algorithm, analysis of convergence, adaptive filter

- Backpropagation (2)

o Multilayer perceptron, function approximation, backpropogation algorithm, architecture, convergence and generalization

- Training issues (2.5)

o Types of training, sample selection, genetic algorithms, ensemble

o Boosting

- Variations of Backpropagation (1.5)

o Drawbacks of backpropogation, heuristic modification, numerical optimization

- Associative Learning (1)

o Simple associative network, unsupervised Hebb rule, decay, simple recognition network (instar), simple recall network (outstar)

- Competitive networks (1.5)