Spring 2011
Small Project 3
Categorization with a simple neural network
Due Friday May 6
Overview
For this project, you will create a simple neural network that learns
to categorize objects based on their features. You will compare its
behavior to that of humans learning the same categories.
Preparation
You will need to create examples that belong to one of two categories.
You may decide to use a variant of the ball
example from class but the choice is up to you. The examples
should be created with the following properties:
- The examples should belong to novel categories for your
participants.
- Each example should be represented with at least 3 features
(e.g. with value of 0 and 1).
- Examples should include relevant and irrelevant features for
categorizing them.
- Examples should have a form presentable to people that have a
clear correspondance to its feature representation.
Human experiment
Before you conduct any experimental trials, work out your protocol.
It should include the following steps:
- Explanation
- Informed consent
- Presentation of training examples
- Testing examples
- Debriefing
Don't forget details such as how you will order the examples and
collect your data. You are strongly encouraged to script key
instructions. Finally, develop all materials that you need for
conducting the experiment (e.g. informed consent statement, script,
stimuli, data logs).
Neural net learning
The web application for creating the neural net is available on the
AI Space Neural Networks
page (click for the Click Here link just below the main title).
Here are some general instructions for creating your neural network
to work with this assignment:
- At the bottom of the canvas, create a row of nodes, each
corresponding to one example feature.
- At the top of the canvas, create an output node (or
multiple output node if you want to assign one category to each
output node).
- Create edges that run from the input nodes to the output nodes.
Note: you are allowed to create intermediate nodes, but they are not
required for this assignment.
- Click on View/Edit Examples in the toolbar to add training
examples and test examples.
- Save your graph and data! (under File)
- Click on the Solve tab to run your neural net. Keep track of
the number of steps. One step corresponds to a training round
using all training examples.
Problems? Post them to the D2L discussion board!
Report
Your report should include the following sections:
- Overview
- Method with human participants
- Results with human experiments
- Method using neural net
- Results with neural net
- Discussion comparing results
- Appendix showing experimental materials (e.g. consent sheet, stimuli)
Using a single-spaced format, the main body of your report may be
as small as three pages.
Submit your report to the corresponding dropbox on D2L.
Grading
This project is worth 20 points. The following grading scheme will be applied when reviewing the reports:
- Coverage (8 points) The report addresses all of the assignment requirements and discussion questions. Specific requirements include:
- Informed consent is addressed.
- The methods are clear enough so that the protocols could
be replicated to achieve similar results.
- The results are a factual account of the outcomes, free
of interpretation.
- The discussion is based on the results. Speculation is
clearly labeled.
- Insight (8 points) The experiments are well motivated.
The discussion uses concepts presented in class (e.g. typicality,
prototypes). Similarities and differences between humans and the
computer model are fully addressed.
- Readability (4 points) The report is well written, easy
to understand and makes helpful use of formatting.