First get the following files from the web 

http://www.cs.toronto.edu/~hinton/csc321/matlab/batchify8.m
http://www.cs.toronto.edu/~hinton/csc321/matlab/doitall23.m
http://www.cs.toronto.edu/~hinton/csc321/matlab/plot23test.m
http://www.cs.toronto.edu/~hinton/csc321/matlab/getscores.m
http://www.cs.toronto.edu/~hinton/csc321/matlab/rbmlearn.m
http://www.cs.toronto.edu/~hinton/csc321/matlab/showrbmweights.m
http://www.cs.toronto.edu/~hinton/csc321/matlab/assignment5data.mat
http://www.cs.toronto.edu/~hinton/csc321/matlab/show.m
http://www.cs.toronto.edu/~hinton/csc321/matlab/getoverlap.m


rbmlearn is a script that implements the learning algorithm for a
restricted Boltzmann machine.  To use it you must set the variable
restart=1;
You must also set values for numhid (the number of hidden units)
and maxepoch (the maximum number of epochs to use for training).

rbmlearn prints out the sum of the squared differences between the
data and the reconstruction of the data from the hidden units.  It
also prints out the weights every 100 epochs.

After you have done
load assignment5data; 
and have set the values for numhid and maxepoch, the script doitall23
will train a model on the dataset train2 and use the model to get
scores for all the images in the datasets test2 and test3.  It will
also train a model on the dataset train3 and get scores from that
model for the two test sets.  Finally it will plot each test case at a
location determined by the scores that the two models give it.

PART 1: (4 points)

You should use doitall23 to investigate how the number of hidden units
and the number of training epochs influence how well the two test
classes can be separated just by looking at their scores under the two
models. Hand in a plot of the test cases for a particular number of
hidden units and training epochs (it does not need to be in
color). Also hand in a description of how you chose the number of
hidden units and training epochs. Your description must be less than
one page.

PART 2: (4 points)

Try training a smaller model and a larger model for each class.  This
will involve modifying the script doitall23.m or writing another
script that calls it and saves the scores that doitall23
produces. Then average the scores of the smaller and larger models
before using plot23test.  Hand in a plot for the smaller model scores,
a plot for the larger model scores, and a plot for the average
scores. Hand in a description that says how you decided on the sizes
of the models and also contains anything interesting that you noticed
about the three plots.  Your description must be less than one page.