Java-Neural-Networks

An old Project of mine to learn AI and neural networks by writing them from scratch based on the fundamental math. This utilizes ejml for the matrices used in the network. This library is not specifically for AI or large scale multiplication like this, so it is CPU bound rather than GPU.

Usage

There are several runners in the project that are used for testing/demonstration purposes only. The pom does not include them in the final build, nor any files they may reference.

If using as a library in another project, do not try to run the runners or MnistImporter(See Future Enhancements below)

FFRunner

Demonstrates the feed forward neural network using the MNist Dataset. Feel free to modify the number of iterations to train more/less

PerceptronRunner

Demonstrates a single perceptron. Perceptrons are capable of learning anything that can map to a linear function.

Sigmoid Runner

Another perceptron demonstration, this time using a sigmoid function. This makes them a better fit for problems with a non-linear space.

Setup

1. Clone repository
2. Setup resources/mnist folder under src/main
3. Locate and download mnist data set. Ensure formatted as instructed in MnistImport.java

Instructions for setup to run yourself are WIP. Stay tuned for updated instructions.

Build

mvn package

This will give you a jar file you can use as a dependency.

Future enhancements

The current build includes the demo package, as well as any files you have in the resources directory of your project. The "Runner" classes in this repo cannot be run without those files, and aren't necessary in the final build for other projects. Moving those to a separate repo and allowing this project purely for the util objects would be better.