Release v4.0.1

Minor changes:

• Added Exponential and Gamma random distributions
• PulseGenerator enhancement (Constant, Uniform, Gaussian or Poisson timing of the leaks)
• Improved performance of the HiddenNeuronPredictors

Release v4.0.0

This release redesigns the core of the RCNet library. It brings an unified hidden neuron and introduces:

• new ability of hidden neurons having analog activations to produce spikes similarly to spiking activations. This also brings a new ability to classify very accurately and quickly using a classical analog ESN reservoir (results look really great)
• dynamic synapse enabling correct communication between hidden neurons having different activations
• unified set of available predictors
• automatic blocking of predictors that give almost constant values. It prevents readout layer learning failure
• simplified and improved xml setup
• Elastic net regression

This fundamental modification also shows that by combining analog and spiking activations, better results can be achieved than each separately. However, this assumption has yet to be verified and will be the main topic of further research.

Release v3.3.1

This release brings minor changes. The most important change is enhanced routing of input fields to Readout Layer (available in SM continuous feeding). Now is possible route only chosen input fields.

Release v3.3.0

This release brings two main changes

• Extreme simplifiaction of StateMachine usage for the developers. Now is not required to care about proper normalization and standardization and then back about denormalization of the input and output data. Both input and output data are expected in the "natural" form and StateMachine does all necessary data operations by itself internally. Building of default text messages containing progress, results and key statistics is now also built-in functionality. Achieved simplification can be seen in DemoConsoleApp source code (SMDemo.cs).
• Introduction of the Ridge Regression and associated RidgeRegrTrainer. Original Linear Regression is now deprecated and reworked to QRD Regression and new associated QRDRegrTrainer. Trainers are using new common method to find good value for the noise and lambda hyperparameters.

Other changes

• Strict split of internal architecture and associated xml setup into the NeuralPreprocessor and ReadoutLayer. Both can be now easily used separately by the developers.
• Incorporated mapper of specific predictors from NeuralPreprocessor to specific ReadoutUnit(s) in ReadoutLayer, including optional alone xml setup ("mapper" element) at the end of the "stateMachineCfg" element.
• Very efficient matrix Inversion and powerful parallelized Mul, Add, Substract, Transpose methods

Release v3.2.1

• Type of input feeding is now independent on the neural task type
• Allowed mixing of classification and forecast tasks on readout layer
• Data handling improvements
• Annotations in RCNetTypes.xsd to make easier writing of xml (typical values of activations parameters)
• Terminology consolidation

Release v3.2.0

Main changes:

• Fixed bug in analog activations
• Dynamic synapse (pre-synaptic short-term-plasticity, post-synaptic decay, signal delay)
• Static synapse (signal delay)
• Improved reservoir's 3D architecture
• Synaptic signal delay driven by Random or Euclidean Distance on reservoir's level
• Removed input spiking neuron
• Virtual inputs based on Const, Random, Mackey-Glass and Sinusoidal signal generators
• Fast Spectral Radius (Power Iteration Method)

Release 3.0.0

@note I apologize, please do not use this release. Release contains bug in all analog activation functions, which makes it unuseable. The latest source code in the repository is now corrected and in the coming days I will publish a new release.

Reservoir architecture enhancement (introduced neuron groups within the pool, better statistics).
Introduced Izhikewich spiking neuron model (one fits all model reproduces spiking and bursting
behavior of known types of cortical neurons).
Consistent transmission signal construction of various neuron types within the reservoir.
Enhanced activation functions parameterization.
Readout layer uses more pesimistic approach for network accuracy evaluation.
Enhanced euclidean distance algorithm.
Enhanced parameterization of augmented states.

Release v2.1.0

Enhanced Spiking neuron models
Ordinary differential equations numerical solver
Vector math

Release v2.0.0

Generalized State Machine

pre-release v1.1.0-beta.4

This pre-release includes a widely parameterized ESN, an independent readout layer and many other components.

Keep in mind:
Future LSM implementation will almost certainly lead to further generalization and merging of ESN with LSM (I hope it will be able to benefit from both approaches). It is almost excluded that the production release will be backwards compatible with this beta release.