Joone

Performance was improved by 50%, thanks to heavy refactoring of the core engine, by adding the single-thread mode. A set of tools was added in order to hide the complexity of the API. A new Image I/O component has been added in order read and write directly from/to image files. A new SoftMax Layer has been added to build neural networks able to resolve 1 of C classification problems. The usability of the property panels has been improved by adopting a new file chooser panel and a visual calendar panel. Several bugs were fixed and the documentation has been updated with the new features.

This release adds support for the Groovy scripting language.
LogarithmicPlugin has been added, to apply a logarithmic transformation
(base e) to input data. "Save as XML" has been added to the GUI Editor,
in order to permit saving a neural network in XML format. A number of
bugs have been fixed, including a problem that prevented SangerSynapse
from learning when in training mode. The inspection panel no longer
shows the biases for Layers for which this doesn't make sense. This
release fixes the lack of the first column when the inspected values
were copied in the clipboard.

This release adds many new features, including the Radial Basis Function, a new Input Connector useful to share the same data source between several layers, a new weights' initializing mechanism, new learning algorithms, etc. A lot of bugs have been fixed and the documentation has been improved.

The Distributed Training Environment (DTE) has
been completely rewritten. The core framework now
has a modular architecture based on plugins,
making it easily expandable. The DTE is easily
configurable in order to define any complex global
optimization technique by using a simple and
powerful parameter file written in XML. The new
DTE is compliant with the latest Sun Jini 2.0.

This version fixes the 'No matching patterns' problem. The
RMSE is not set to 0.0 when a network is deserialized. A
NullPointerException was fixed in the
YahooFinanceInputSynapse when no input data is present.
The BatchLearner has been fixed in order to use the
momentum parameter.