AI Lab #8

CS63 Fall 2005
Lab 8: Neural Networks
Due: Monday, November 14 by 11am

Introduction
General guidelines
Option 1: Images
Option 2: Robot control
What to hand in

INTRODUCTION

This lab gives you the chance to formulate a machine learning problem and use a neural network to solve it. First you will learn about conx, the Pyro module that implements neural networks. Then you may choose one of two options for your neural network experiments: work with a data set containing images of people's faces, or try a robot control task.

You may work with a partner on this lab.

GENERAL GUIDELINES

Before using conx you'll need to set the PYTHONPATH variable in your shell. You can find out which shell you are using by doing: finger username.

For bash shell, add the following to your .bashrc file:
```
export PYTHONPATH=/usr/local:/usr/local/pyrobot
```
For tcsh shell, add the following to your .cshrc file:
```
setenv PYTHONPATH '/usr/local:/usr/local/pyrobot'
```

Any new shell window you open should now be able to find the conx module.

Read through PyroModuleNeuralNetworks. Focus on the first two sections of this module.

The source code of conx is viewable at:
```
/usr/local/pyrobot/brain/conx.py
```
You may also want to read about the details of particular conx methods at ConxImplementationDetails

For a review of how to set up a well designed neural network experiment, read the handout from Mitchell's Machine Learning.

When designing your experiment be sure to use a training set and a validation set so as to avoid overfitting (for an example of how to do this, look at the NNvoting.py example we discussed in class).

Define your learning task and then through trial and error determine the best network architecture for accomplishing it. We'd like the smallest hidden layer size for which the task can be consistently learned.

After training is complete, you should use some form of analysis to examine the network's solution. For the image data you can view the weights from the input to hidden layer as an image. In conx, you can get the input to hidden weights by using the following Network class method:
```
getWeights('input', 'hidden')
```
For either task, you can use cluster analysis on the hidden layer activations. For a robot task, principal components analysis might also be useful.

OPTION 1: IMAGES

The directory /home/meeden/public/cs63/faces_4/ contains 625 images stored in PGM format. You can view one of these images using the xv command. Each file is named according to the following convention:

userid_pose_expression_eyes_scale.pgm

userid is the user id of the person in the image. This field has 20 values: ani2, at33, boland, bpm, ch4f, cheyer, choon, danieln, glickman, karyadi, kawamura, kk49, megak, mitchell, night, phoebe, saavik, steffi, sz24, and tammo.
pose is the head position of the person, and this field has 4 values: straight, left, right, and up.
expression is the facial expression of the person, and this field has 4 values: neutral, happy, sad, and angry.
eyes: is the eye state of the person, and this field has 2 values: open and sunglasses.
scale: is the scale of the image. All of the images are of type 4 which indicates a quarter-resolution image (32 x 30).

I have written some python functions to help you process these image files and also to help you visualize the weights as an image as shown in the Mitchell handout.

There are many possible classification tasks you could perform on these images, such as determining head position or whether someone has sunglasses on. Some of these classification tasks, such as whether someone is wearing sunglasses, are relatively easy for a neural network to learn. You may want to try several tasks of varying difficulty and compare the results.

OPTION 2: ROBOT CONTROL

Read through Robot learning using neural networks.

Formulate a robot learning task. You'll need to think carefully about how to represent the inputs and outputs. Will you use offline or online training? Beware that online training can lead to the problem of catasrophic forgetting. This occurs when the robot is exposed to many similar situations in succession and becomes overtrained on this situation, forgetting how to respond to other situations.

Learning in a robot domain can take an extremely long time. If you plan on trying this option you must start early to be able to complete it in time. Therefore, by Monday you must email me with a description of your task and specifically how you plan to implement it in a neural network.

HAND IN

Use cs63handin to turn in a tar file of an entire directory that should contain all of the code needed to run your experiments, as well as examples of your post-training analysis, and a text file. The text file should explain the experiment in enough detail so that it could be easily reproduced. You should also discuss the results and refer to interesting features found in your post-training analysis.

To tar up your directory do the following:

tar -cf nnet.tar directoryName