HandSignClassifier

A neural network built in Java from scratch for recognizing numerical hand signs

Overview

The repository consists of three components:

1. ABCDImageNetwork: A four-layer feedforward neural network implemented via backpropagation with configurable settings:

   1. Training/running mode
   2. Weight randomization/file input
   3. Weight file output
   4. Data file input
   5. Network layer sizes
   6. Training parameters: $\lambda$, $E_{max}$, and $n_{iterations}$

2. imageProcessing: Converted data files for hand sign images.

3. handImages: Raw numerical hand sign images.

This project was created for ATCS: Neural Nets at The Harker School, taught by Dr. Nelson.

Usage

Dependencies

Standard Java libraries.

Running

1. Edit the control file, which has four arguments:
   1. doTrainNotRun: Whether to train or run (boolean)
   2. networkConfigurationFilename: The network configuration filename (string)
      • This network configuration file contains values for network parameters.
   3. inputSetFilename: The input set filename to run the network on (string)
      • This input set file in turn contains an ordered list of individual input member files.
   4. targetSetFilename: The target set filename to compare with the network's outputs (string)
      • This output set file in turn contains an ordered list of individual target member files.
2. Run ABCDNetworkTester.java from the terminal with the control filename as the first argument.
3. The program will create and run the network based on the provided settings.
4. Once finished training/running, the program will print the network specifications and a comparison table of network outputs and target outputs to the console.
   1. If specified, the network will also save weights to the specified weight output file.

Help

See the repository's control, network configuration, input set, and target set files for examples.