Past Participant Survey
Challenge Team Interim Report
Team Number: 047
School NameCimarron High School
Area of Science:Computer Science
Project Title:Neural Networks
A Neural Network is computer architecture modeled upon a brain's
interconnected system of neurons. Most neural networks are simulations run
on conventional computers. They imitate the brain's ability to sort out
patterns and learn from trial and error. Neural Networks are well suited
for pattern recognition, foreign language translation, process control,
medical data interpretation, and parallel processing. Unlike comparing
numbers in standard programs, neural nets can learn to recognize near
Neural Networks can compute any computable function. They can do
everything a normal digital computer can do, perhaps more. They are
especially useful for classification and mapping problems.
Often the purpose of using a neural net is to generalize. In order
for a network to have good generalization three conditions are necessary.
The first is that the inputs to the network contain sufficient information
pertaining to the subject, so that there exists a mathematical function
relating correct outputs to inputs with the desired accuraccy. A network
cannot learn a nonexistent function.
The second condition is that the function being learned is smooth.
A small change in the inputs should produce a small change in the outputs.
For continuous inputs and targets, smoothness of the function implies
continuity and restrictions over most of the input space. It is more
likely to get better generalization with realistic sample sizes if the
classification boundaries are smoother.
The third condition for good generalization is that the training
case be a sufficiently large and representative subset. The importance of
this is related to the fact that there are two different types of
generalization: interpolation and extrapolation. Interpolation applies to
cases that are more or less surrounded by nearby training cases;
everything else is considered extrapolation. Interpolation can often be
done reliably, unlike extrapolation which is often unreliable. It is
important to have sufficient training data to avoid the need for
We plan to modify an existing type of neural net program and train
it to recognize the differennt patterns for different brands of bottom
soles of shoes. The test will be for the net to identify a sole pattern it
has not seen as to which manufacturer it is from. This pattern should have
all the characteristics mentioned for patterns that neural nets can
- Jeffrey Raloff
- Dean Bernadone