Search for Science Fair Projects

1000 Science Fair Projects with Complete Instructions

Computer Science Science Fair Project

Neural Network Digit Recognition

Hard
Neural Network Digit Recognition | Science Fair Projects | STEM Projects
How well can a program learn to read handwritten numbers from scratch? You build a neural network (a program that learns by adjusting connection weights) with three layers. Then you train it on twenty examples of each digit from 0 to 9. The network uses back propagation (a method that corrects mistakes layer by layer) over 30,000 training rounds. After training, it identifies the practice digits with 90% accuracy. Then you test it on new handwritten digits from people who never saw the training set. The network recognizes these fresh digits with 83% accuracy. It adapts to unfamiliar handwriting and still reads the numbers correctly.

Hypothesis

The hypothesis is that if a neural network is trained to identify decimal digits, then it will be able to identify the digits with a 75% accuracy.

Method & Materials

You will design and code a neural network to identify decimal digits using Visual Basic and Excel. You will train the neural network with twenty examples of each digit, for a total number of 200 trials. You will use the back propagation algorithm to train the neural network, iterating 30,000 times and achieving a 0.01 mean square error.
You will need a computer, Visual Basic, Excel, and data for the neural network.

MEL Math — monthly math & computer science experiment kits delivered to your door. (Affiliate link)

See whats included

Results

The results of this project showed that the neural network was able to identify the training set of digits with a 90% accuracy and the experiment data with an 83% accuracy. This proves that the hypothesis is valid.

Why do this project?

This science project is unique because it uses a neural network to identify decimal digits using Visual Basic and Excel.

Also Consider

Experiment variations to consider include using different numbers of iterations for the training data and using different learning rates for the weights.

Full project details

Additional information and source material for this project are available below.
Share this Science Project:

Related Science Fair Project Ideas

CPU Speed vs. RAM Size in Computer Performance
Swap the CPU and RAM in a desktop computer and time the difference to find out which upgrade matters more.
Hard
Programming the Diffusion Equation
Write a program in Scheme that solves differential equations and use it to model how atoms spread through solids.
Hard
Billiard Ball Angles and Corner Pockets
Simulate 6,300 billiard ball paths and discover that a single property of the launch angle predicts whether the ball reaches a corner.
Hard
Share this Science Project: