Artificial Neural Networks

Course Nos. ECE.09.560 (Grad), ECE.09.454 (Senior Elective)

Fall 2008

Lab Project 2: Multilayer Perceptrons

Objective

In this project, you will develop techniques for designing multilayer perceptron models for solving complex function approximation and classification problems. This project has four parts. In Part 1, you will exercise your MLP model to solve canonical logic-gate problems. In Part 2, you will attempt to use your MLP model for performing function approximation. In Parts 3, and 4 you will solve pattern recognition problems.

Pre-Lab Exercise

Multilayer Perceptrons are modeled in Matlab in a 4-step process:

The function newff creates a feedforward MLP neural network
The function net.trainParam is used specify the training parameters to the MLP
The function train is used to train the MLP
The function sim is used to simulate/test the MLP

help newff

See classroom demo Multilayer Percepton - Backprop Training.

Experiment with different number of training passes, input-output configurations, etc.

Part 1

In this part, you are required to demonstrate the capability of an MLP to model the XOR logic gate. Generate performance curves/surfacess for these MLP-models as the inputs vary continuously from 0.0 to 1.0.

Part 2

In this part you are required to demonstrate the capability of an MLP to approximate the function f(t) = sin(t)*exp(-t/20); 0 < t < 50 Experiment with varying the number and location of training data points, number of hidden layers and nodes, etc. What are your conclusions?

Part 3

Repeat Part 2 of Lab Project 1 by using a multi-layer perceptron model to separate classes in the Iris database of the UCI Machine Learning Repository: http://www.ics.uci.edu/~mlearn/MLRepository.html

Part 4

In this part, you will repeat Part 3 of Lab Project 1 and construct an MLP to identify a subset (of your choice) of the students in the class. Jpeg image files of student pictures can be downloaded from classpics. You would need the following Matlab functions to convert the image files into a vector suitable for neural network input:

imread to read the image,
rgb2gray to convert the image to grayscale,
mat2gray to convert the image into a matrix in the range 0-1,
imresize to shrink the image,
imshow to display the image,
reshape to convert the matrix into a vector.

(do help with all of these functions for more details).

As part of your neural network design process, you will experiment with

Choosing appropriate training and test data
Data preprocessing - feeding raw image vectors vs. truncated FFT's or DCT's of the images.
Number and type of output vectors
Network architectures (number of hidden nodes and layers)
Training algorithms and strategies
Images corrupted with noise (use the imnoise function). Identify the SNR for network failure. Not familiar with SNR concepts? See ECOMMS Class Lab Project 1 Pre-Lab Lecture and ECOMMS Class Lab Project 1.

Tabulate your percentage correct classification results for each of these runs. Be aware that as in any real neural network, you will have misclassifications. You are required to draw conclusions from this study as you develop the most "optimal" MLP model for this problem.

Your laboratory report (individual report is required of each student) should be in the usual format.

Instructor Schedule Textbook Links Grading Homepage