605.649 — Introduction to Machine Learning Programming Project #5


5/5 - (2 votes)

605.649 — Introduction to Machine Learning
Programming Project #5
The purpose of this assignment is to give you experience developing one of the main algorithms for
training feedforward neural networks—backpropagation. In this project, you will evaluate the performance
of a feedforward network trained with backprop. You will also use the same five datasets that you used from
Project 1 and Project 4 from the UCI Machine Learning Repository, namely:
1. Breast Cancer —
This breast cancer databases was obtained from the University of Wisconsin Hospitals, Madison from
Dr. William H. Wolberg.
2. Glass —
The study of classification of types of glass was motivated by criminological investigation.
3. Iris —
The data set contains 3 classes of 50 instances each, where each class refers to a type of iris plant.
4. Soybean (small) —
A small subset of the original soybean database.
5. Vote —
This data set includes votes for each of the U.S. House of Representatives Congressmen on the 16 key
votes identified by the Congressional Quarterly Almanac.
When using these data sets, be careful of some issues.
1. Some of the data sets have missing attribute values. When this occurs in low numbers, you may simply
edit the corresponding values out of the data sets. For more occurrences, you should do some kind of
“data imputation” where, basically, you generate a value of some kind. This can be purely random, or
it can be sampled according to the conditional probability of the values occurring, given the underlying
class for that example. The choice is yours, but be sure to document your choice.
2. Most of attributes in the various data sets are either multi-value discrete (categorical) or real-valued.
You will need to deal with this in some way. You should apply what is called “one-hot coding” where
you create a separate Boolean attribute for each value. For the continuous attributes, you may use
one-hot-coding if you wish, but there is actually a better way. Specifically, it is recommended that you
normalize them first to be in the range −1 to +1. (If you want to normalize to be in the range 0 to 1,
that’s fine. Just be consistent.)
For this project, the following steps are required:
• Download the five (5) data sets from the UCI Machine Learning repository. You can find this repository
at All of the specific URLs are also provided above.
• Pre-process each data set as necessary to handle missing data and non-Boolean data (both classes and
• Implement backpropagation for training feedforward neural networks. You may choose whether or
not you wish to use the logistic activation function or the hyperbolic tangent activation function.
Remember that this choice affects the update rules because of having different derivatives. It is also
your choice whether or not you use momentum.
• Run your algorithms on each of the data sets. These runs should be done with 5-fold cross-validation
so you can compare your results statistically. You can use classification error or mean squared error
(as appropriate) for your loss function.
• Run your feedforward network trained with backpropagation on each of the data sets. These runs
should output the classifications on all of the test examples. If you are doing cross-validation, just
output classifications for one fold each. Run your experiments with no hidden layers, one hidden layer,
and two hidden layers. It is up to you to tune the number of hidden nodes per layer, and be sure to
explain how you did the tuning.
• Write a very brief paper that incorporates the following elements, summarizing the results of your
experiments. Your paper is required to be at least 5 pages and no more than 10 pages using the JMLR
format You can find templates for this format at The
format is also available within Overleaf.
1. Title and author name
2. Problem statement, including hypothesis, projecting how you expect each algorithm to perform
3. Brief description of your experimental approach, including any assumptions made with your algorithms
4. Presentation of the results of your experiments
5. A discussion of the behavior of your algorithms, combined with any conclusions you can draw
6. Summary
7. References (Only required if you use a resource other than the course content.)
• Submit your fully documented code, the video demonstrating the running of your programs, and your
• For the video, the following constitute minimal requirements that must be satisfied:
– The video is to be no longer than 5 minutes long.
– The video should be provided in mp4 format. Alternatively, it can be uploaded to a streaming
service such as YouTube with a link provided.
– Fast forwarding is permitted through long computational cycles. Fast forwarding is not permitted
whenever there is a voice-over or when results are being presented.
– Be sure to provide verbal commentary or explanation on all of the elements you are demonstrating.
– Provide sample outputs from one test set showing performance on your feedforward networks.
Show results for each of the cases where you have no hidden layers, one hidden layer, and two
hidden layers.
– Show a sample model for the smallest of each of your three neural network types (i.e., zero hidden
layer, one hidden layer, two hidden layers). This will consist of showing the weight matrices with
the inputs/outputs of the layer labeled in some way.
– Demonstrate and explain how an example is propagated through a two hidden layer network. Be
sure to show the activations at each layer being calculated correctly.
– Demonstrate the weight updates occurring on a two-layer network for each of the layers.
– Demonstrate the gradient calculation at the output for any one of your networks.
– Show the average performance over the five folds for one of the data sets for each of the three
types of networks (i.e., zero hidden layer, one hidden layer, two hidden layers).
Your grade will be broken down as follows:
• Code structure – 10%
• Code documentation/commenting – 10%
• Proper functioning of your code, as illustrated by a 5 minute video – 30%
• Summary paper – 50%

Scroll to Top