NeuroBox News

New: NeuroBox Repository, a Network Structure and Training Status Storage Manager, based on ADO.NET typed DataSets allowing to easily import/export to XML Files and various relational databases.

The NeuroBox Library is now splitted in 3 smaller projects:
NeuroBox.dll: Neural Network Core
NeuroBox.PatternMatching.dll: Pattern Matching Building Block
NeuroBox.Repository.dll: Network IO Manager

There's a new (simple) demo available: A GUI logic operator training demo, working directly with the network classes (does not use building blocks).

NeuroBox introduces transparent bias neurons in this release.

Bias neurons (aka on-neurons) help braking down network
symetries by introducing a constant bias network input
for each neuron, adjustable with trainable weights
as know from all other neurons.

Bias neurons are now included in the 'optimized'
configuration subsets, but are disabled by default.

To enable bias neurons:

network.Configuration.UseBiasNeurons = true;... read more

This is a minor release:

- New small (but growing) Demo Collection
-> Simple Training Demo
- Important BackPropagation gradient bug fixed.
-> Now learning much faster
- Bug in Success Evaluation in the Pattern Matching
Building Block fixed.
- Pattern Matching Demo now displays activity instead
of output of the output neurons
- Some missing network methods added (unbound scenario)
- Preparation for batch based training (in addition
to the current incremental alogrithm)
- New network structure factory, to build a network
using the structure defined in two arrays,
as requested in the sourceforge trackers.

- Some minor fixes.
- PropertyGrid Support for configuration class.
- Some new properties and methods, some moved
(eg. a Neuron now can create synapses itself, instead
of the roundtrip to the network class needed in v2.0)
- Improved and new Network Training Optimizations.
- New Event driven architecture for better integration.
- New Building Blocks (-> BuildingBlocks Namespace):
- GridPatternMatching
- Pattern management
- network structure factories for much smarter selective
synapse creation (instead of the 'connect all to all'
approach)
- etc.... read more

- cleaner and simpler architecture
- much better layer management
- easier to use
- commented, reference available
(on the project's website also in html format)

NeuroBox is now compatible to the .NET Framework RTM (Final). The Project/Solution files are upgraded to VisualStudio.NET final. Also, the binary files are compiled for the Framework Version v1.0.3705.

As you've requested in the survey I'm building a demo client right now, based on this release. I'll upload as soon as it's available ...

btw: In my robotic application xbot here at sourceforge there's a (pseudo) visual module using neurobox implemented for designing and training small feedforward networks....

The development of a new neuronal network architecture implementation for NeuroBox has been started. The asynchronous, differential net inherits the current network classes but is a better representation of real biological processes. It can be used especially for linguistic networks, knowledge management and controlling system and other asnyc algorithms.

The first public release of the NeuroBox Library, wich generates, propagates and trains neuronal feedforward networks using the .NET Framework, is now available. It supports basic features like propagation and backpropagation with flatspot elimination, momentum term etc.

The coming major release, 2.0, will also support backpercolation.