Assignments
Due
dates
 All assignments will be due on Wednesdays at the
start of class.
 Assignments will be posted on this web
page one week before they are due.
Marking scheme and lateness penalties
 Each of the 3 assignments will be worth 15% of the final grade.
 Except in the case of an official Student Medical
Certificate, assignments that are submitted late will have their
score reduced by 25% per day or part
of a day (so after
4 days they will get zero however good they are). The time past the
deadline will be rounded UP to an integer number of days.
Collaboration Policy for Assignments
 You are expected to work on the assignments by yourself.
You should not discuss them with anyone except the tutors or the
instructor. The report you hand in should be entirely your own work.
 Assignment 1
assoc.m
This involves answering some questions about generalization,
understanding a short matlab program, and
implementing the backpropagation
algorithm to learn to recognize handwritten digits.

Assignment 2
You will need to download the following matlab files to use the
modified version of minimize.m. Do not use the version available from
Rasmussen's webpage.
minimize.m
checkgrad.m
ECG.m
runme.m
toydata.mat
toydata_test.mat
Files ECG.m and runme.m demonstrate how minimize.m can be used to
train a neural network with 3 hidden layers. runme.m is the driver
routine that alternates between calling minimize.m and computing
training and validation errors. ECG.m is the function passed to
minimize.m that computes the gradients and outputs for the neural
network. Note that since the mixturefitting code you need to write
will be much simpler than the code in ECG.m and runme.m, your time
might be better spent reading the documentation for minimize.m rather
than trying to understand ECG.m and runme.m in minute detail.

Assignment 3
Many people found assignment 3 difficult. I am therefore
making it faster to do the empirical part by reducing the amount of
training data and I will explain the learning procedure for RBM's in
much more detail in the lecture on Nov 5. The deadline for assignment
3 is now wed Nov 12 at 1.00pm. Please tell anyone else you know in
the class. The reduced training data will be just the first two
minibatches of the current training data so there will only be 20
examples of each of the 5 digit classes.
If you have already finished assignment 3 you are welcome to hand in
it in as is.
Andriy Mnih will be grading assignment 3. He will have his usual office hours from 3.00
to 4.00 on thursday nov 6, but he will NOT have office hours on monday Nov 10.