Jo Dene neural network library in Java, as an open source download under the Apache public license.
The Jo Dene neural network library is copyright 2004, Paul Hoehne
For more information contact Paul Hoehne
For more information about the apache license contact Apache
05/23/04 - Refactored the package structure. Now com.penguinwerks.jodene is the root package for interfaces and base classes. Other packages contain implementations of interfaces and usefull utilities. Also updated documentation.
05/18/04 - Documentation updates.
05/17/04 - Added pocket algorithm with some testing. Note that now there is a dependancy on log4j, which is available at the Apache website. Over time all the infamous, quick and dirty println's will be replaced with logging commands. I know that documentation has not been updated in a long time. That will be the next priority for the next update.
05/11/04 - Tested tanh activation function and finished an end-to-end training test. The tanh function requires a much lower learning rate than the sigmoid function. In fact, it's less than 1/4 of the learning rate for the sigmoid function, even though the derivative of the tanh function is only four times as large as the derivative for the sigmoid.
05/09/04 - Added but not tested tanh scaling activation functions. Fixed the NetworkBuilder class (but not complete). Unfortunately not much accomplished due to house painting and a M$ windows crash while installing SuSE 9.1. Which, BTW, is a solid distribution, but I find the Apache/PHP included within the distro to be less than satisfying (compiled without PEAR, but ships with all the PEAR libraries).
05/04/04 - Refactored error calculators, listeners, an activations into separate packages. The interfaces they implement are still in the networks package, but the implementation are in separate packages to keep down the size of the networks package. Added an input data manager. Added factories for activation functions.
05/03/04 - Adding managers to store and retrieve networks, training sets, etc. In some future version they may even be remoted. Also, tested serializtion and de-serialization of networks. Networks now train on Iris data set from the UCI machine learning repository.
04/25/04 - Refactored code. Now the trainer uses an error manager than can track a variety of errors during training.
04/21/04 - Named after my darling wife, the Jo Dene Neural Network library is a simple feed forward/backprop network library implemented in Java. Documentation is available.