JNeuralNet

Best neural network resource

2009-06-28T00:28:00.000-07:00

Today i happened to find the best neural network resource (according to me). It almost has every thing from sample data sets to awesome tutorials. Check it out

Introducing JNNHelperNB 1.0b

2009-06-14T02:32:00.000-07:00

Check the new JNNHelperNB here, a GUI helper tool based on Netbeans Platform 6.7. It is completely ported and is now easier to maintain and develop. We are planning to include an auto update feature and an online gallery of the contributed plugins for the available SPI's. This tool can be checked out from subversion repository using the following command:

svn checkout https://jneuralnet.dev.java.net/svn/jneuralnet/trunk JNNHelperNB --username guest
(leave password field blank when prompted)

Also, the visual network view is being replaced by widgets of Visual API. It'll be included along with property sheets for editing neuron properties in the next milestone release.

As of June 10th 2009, JNNHelper tool development is obsolete. Use the project issue tracker to add your wish list.

The Netbeans experience

2009-06-10T07:26:00.000-07:00

It has only been a week since we started porting our application 'JNNHelper' to netbeans platform and we are already seeing its benefits.

The lookup API is doing an awesome job in decoupling module dependencies which was otherwise going haywire. We are glad that there is so much documentation available out there...moreover Netbeans mailing lists rock, most of our questions actually get answered!

We can only hope that out new application 'JNNHelperNB' will be out there by the end of this month. We are planning to add many new features in this release including the use of Visual API for cooler visualization and fine grained customization support using property sheets.

Temperature forecasting sample...

2009-05-30T06:55:00.000-07:00

Finally we managed to test the application for a real world situation. Check out the temperature forecasting folder in documents and downloads section of JNeuralNet home. It has a detailed project report on how such an application was built (along with the necessary mathematical background, not required..optional).

It marks a big achievement as the prediction has an average of 83% success rate on unknown data sets. More over using the JNNHelper tool this has been acheived in 5 mins (including training and analysis).

Future plans for JNeuralNet

2009-05-27T04:08:00.000-07:00

The following are the changes planned for the project.

We were planning to cluster out JNNHelper tool as a sub project of JNeuralNet.

Move project to subversion for nightly builds. Only the milestone releases will be available for direct download.

Enhancements include:

Add project save and load.

Neural network check pointing while training (thanks to chandan for this idea)

Wizard to create new algorithms

Inbuilt editor for creating pluggable components.

On online gallery of pluggable components, these will further be available as online updates for the GUI tool.

Property sheets for fine grained customization of the neural network.

The last four enhancements are the reason we are planning to move onto Netbeans platform. It has extensive documentation at Netbeans Platform Home. I personally found these screencasts to be excellent, a must see if you want to learn the netbeans platform.

These enhancements will be added in the next milestone release (hopefully). Suggestions and comments are welcome.

We are also planning to include screen casts for tool usage, plugin development and demo applications using Wink, contributors are more than welcome.

All New JNeuralNet

2009-05-27T03:46:00.000-07:00

Its been a long time since i have blogged...was busy with acads and other stuff. Any ways, there have been several exciting updates to the JNeuralNet project.

Its now possible to create pluggable extensions of activation function, learning algorithm, cost functions, input and output preprocessors and datastores. For more information on these spi's please visit the project home

An all new JNNHelper tool 1.2 featuring stable network builder, data builder, trainer and analyzer.

We have also included a temperature forecasting application as a demo using the helper tool (included a detailed project report)

Fully updated javadocs (finally) , architecture docs and other support documentation.

Test drive the application, criticisms and suggestions are more than welcome. please use the projects mailing lists for that.

Enhancements Planned

2009-02-02T09:33:00.000-08:00

While trying to build an application using JNeuralNetwork i've realized that a few things were missing.

1) Activation function choice - The default sigmoid activation is not a very good choice if the output range is other than (0, 1). Hence an interface is planned and its default inplementations of popular types are planned.

2) Plugin architecture for learning algo's and activation functions using java service discovery. This feature is planned to be integrated into the helper tool. Also an editor is provided to create new types, generate plugins to aid in development.

3) XML based save and load for neural networks.

4) An option of JAR export is planned. This is needed incase plugins are used. The exported jars are to be included in the library of a larger application using the neural network. One simply has to call MyNN.getOutput(input) to get the prediction...No need to know ANY of the base API's.

The above features are under development and will be out there by the end of february...

Features

2008-11-17T21:10:00.000-08:00

The following are the planned features for the GUI helper tool. Any suggestions and improvements are highly appreciated. We’ll try to implement any new features you may suggest.

1) An intuitive GUI with five tabs

Network designer
Training Data
Validation Data
Training analysis
Performance analyzer

2) The minimum data you’ll have to provide is the training data or the data set. The remainder of the configuration, i.e. network design, choice of validation data and choice of training algorithm to be used can be configured optimally by an ‘Auto Configure’ button.

3) Training and performance analysis shows a graphical display (thanks to JFreeChart project) conveying the necessary information to the user. The necessary information can be logged (uses log4j) for later analysis and the graphs can be saved in any of the image formats.

4) Can export and import the neural network. At this point we’re at a bit of confusion. Should the save be exported in XML format or should serialization be used? If XML is used what could be the potential advantages? Please comment on this issue. Current object serialization is used.

Take a look at the usage doc, UML model we’ve prepared. Is it overly complex?, can it be simplified further?? We look forward for your comments on this issue.

Why another framework?

2008-11-17T21:07:00.000-08:00

There are many frameworks that exist to simplify the Neural Network creation and training. My personal favorite is the JOONE engine. It provides an intuitive GUI and a well defined engine that is capable of creating modular neural networks (combination of one or more types).

When I first used the JOONE engine, I was overwhelmed with possibilities. It had too many menus and advanced functionalities and took me a while to understand the functionality. Essentially there are five reasons why we chose to design a new framework.

1) 90% of neural network related problems are focused around supervised networks; the back propagation network is the most popular amongst them. Hence the idea of creating a framework exclusively for supervised BP networks.

2) I have seen many students who were interested in neural networks, but backed off due to the underlying mathematical complexity. The aim of the project is to abstract the mathematical complexity while providing the necessary flexibility for an advanced user to fiddle with various parameters for research.

3) As JOONE was designed by keeping a broader perspective in mind, instead of reusing it, we’ve defined a simpler framework exclusively for the back propagation network.

4) We have planned to include a basic GUI helper tool and an advanced GUI tool targeting both the beginner and advanced counterparts.

5) Most of the people are overwhelmed by number of adjustable parameters to design that perfect neural network tailored for your needs. For this we’ll soon be providing and ‘auto configure’ option that sets the most optimal architecture for a given data set by using genetic algorithms.

In summary, emphasis is laid on the simplicity of usage and at the same time giving it sufficient flexibility for customization. Moreover, creating a neural network is as simple as following 5 steps.