Release Notes: This version introduces early stopping and colour-encoded error displays. It uses double-precision floating point numbers like version 0.1 and 0.2. Some bugs were also fixed.
Release Notes: A bug in the WAV/ASCII save procedure has been fixed.
Release Notes: This one can read and write Ascii and WAV files, some bugs were removed, and the interface is more comfortable.
Release Notes: An option to connect neurons with different connection types simultaneously was added. An option to upload algorithm settings from a file directly to a selected connection was added. A statusbar was added.
Release Notes: This version includes algorithm save/load functionality plus two new activation functions: the 'Mexican Hat' and the 'Gaussian Derivative'.
Release Notes: A momentum parameter has been introduced to speed up the training algorithm. Furthermore, a new update option has been introduced to discard potentially harmful updates.
Release Notes: This version adds the option to use the hyperbolic tangents or the Gaussian bell for gate activation.
Release Notes: This version is able to process more than one set of signals. Every set of signals is treated as a different pattern. All patterns where at least one target signal is provided can be trained. Furthermore, RTRL now allows post-update gradient recalculations to improve precision.
Release Notes: The LSTM neuron model has been extended to 'Forget Gates'.
Release Notes: This release has been extended to LSTM (Long Short-Term Memory) neurons to overcome RTRL limitations. It is about seven times faster than 0.4.