Ubiquitous Neural Networks

Ubiquitous Neural Networks is a framework and collection of neural network models that stemmed from my PhD Thesis. The original code was completely refactored and is now made available through this project.

It allows you to easily build and train neural networks in Java; you can also extend the core functionality for particular use-cases.

It includes a variety of features and visualizations to help you create powerful and accurate models for your data.

🚧 Check out the Wiki pages.

Features

Models

This project implements three types of neural network models and variants:

SelfOrganizingMap (SOM): a type of neural network that projects high-dimensional data onto a 2d map, maintaining input-space topological relationships. Useful for exploratory cluster analysis and, possibly, classification.
- BasicSOM - Classical and Batch algorithms are available;
- For streaming data, i.e., variants that can learn incrementally over time, estimating its learning parameters on-the-fly:
  - UbiSOM - The Ubiquitous Self-Organizing Map (a contribution from my thesis);
  - PLSOM - The Parameterless Self-Organizing Map;
  - DSOM - The Dynamic Self-Organizing Map.
All SOM models allow setting the lattice type (hexagonal or rectangular) and the metric distance to use (euclidean or manhattan); but you can easily create your own lattices and metric distances.
StreamART2A (a contribution from my thesis): a neural network architecture, based on Adaptive Resonance Theory, that uses a process of competitive learning and resonance to summarize input patterns in streaming data.
MLPNetwork (Multilayer Perceptron): a feedforward neural network architecture consisting of multiple layers of nodes that can approximate any function with enough hidden units. Useful for classification and regression.

You can create deep learning models and comes with several activation functions, e.g., linear, ReLU, sigmoid and tanh.

Data and preprocessing

Data is imported through the Dataset class. Data set files must have a yaml header describing it - see examples in the datasets folder;
There are two implementations of DatasetNormalization, namely MinMaxNormalization and MeanNormalization. You can add others;
The class DatasetTrainSplit, as the name suggests, allows you to easily split your data for training and testing;
The library comes with an implementation of PCA data projection.

Visualizations

Contains different types of visualizations for all available models and a simple plotting class.

Self-Organizing Maps

U-matrix
Component planes
Hit-Map
Spatial visualization of input data and lattice
Clustering of prototypes
...

StreamART2A

Spatial visualization of input data and generated micro-categories

MLP Networks

Network architecture, weights and bias

Documentation

Installation

Binaries and dependency information for Maven, Gradle and others can be found at http://search.maven.org.

Example for Maven:

<dependency>
    <groupId>com.brunomnsilva</groupId>
    <artifactId>neuralnetworks</artifactId>
    <version>x.y.z</version>
</dependency>

You need Java 9 or later.

Some transient dependencies are used, namely:

API Reference

Usage and examples

🚧 Check out the Wiki pages.

⚠️ Until they are populated with examples, you can check some usage examples in the neuralnetworks-examples folder.

Citing

🚧

Contributing

You can fork the repository and submit a pull request. Pull requests should adhere to the existing naming and Javadoc conventions.

License

This project is licensed under the MIT License - see the LICENSE file for details. All derivative work should include this license.

Authors

Original author: Bruno Silva - (GitHub page) | (Personal page)

Contributors

(no others, at the moment)

I hope you find Ubiquitous Neural Networks useful and look forward to seeing the projects you create with it!

brunomnsilva / UbiquitousNeuralNetworks