![]() |
|
||
|
Development Notes 01/16/2007 - Fixed a bug that could prevent the network thread from restarting after being shut down earlier. Development on the neural net hasn't seen much action lately as I am instead focusing on rewriting the E-Mail server. 11/08/2006 - A new interface is available for driving input into neural nets. From the Net menu, you can insert input pads. These pads of 3x3 to 16x16 squares let you define patterns by clicking squares on and off. Drag a connection from a square to a neurone, and that square will send a pulse each cycle to its connected neurone. This allows you easily to see how different patterns affect your net! 09/26/2006 - The neural net simulator now has partial support for accessibility through MSAA! Although I don't have access to braille embossers or anything else of the sort, I have tested the accessibility using Microsoft's free Accessibility Explorer, and the neurones and axon terminals show up as separate accessible objects. I doubt the support is currently enough to help someone with disabilities, but I'll keep working on it. 07/31/2006 - Added the SendTextStream and PromptTextStream methods to the scripted neurones. This means you can stream text (one bit per cycle) into nets for various purposes. 07/16/2006 - Now you can watch the values of neurones on a chart! Right click a neurone and click "Chart Value." This will give you a popup line graph chart. 07/10/2006 - The simulation has partial Kohonen neurone support! It lacks a lot, but a big part of the necessary architecture is in place. In the case of Kohonen neurones, the architecture now supports grouping for special neurones. The grouping is what allows "competitive learning," although this is not yet supported. However, a Kohonen neurone group will pick one winner per cycle, and only that neurone from the group will fire. 04/29/2006 - Scripted neurones can programmatically manipulate trainable neurones using a new dispatch interface. Call Net.GetTrainables to get the the interface. See the Scripting.txt file for details. 04/28/2006 - Added a new type of trainable neurone. The new type only fires if its accumulated value meets or exceeds its threshold each cycle. If the value does not meet the threshold, it is reset. Recently, I discovered that it is with the "greater than or equal" neurones that XOR style circuits can be constructed, so I decided that property might be useful for the trainable neurones. 04/17/2006 - Changed the format of the SNN files to support weighted connections. Weighted connections bring the simulator much closer to what you would expect to find in a neural net simulation program. The weights allow a single neurone to send pulses of different strengths to multiple neurones. In addition to weighted connections, all calculations now use floating point numbers, and the connection weights can be negative (inhibiting neurones). Since the file format has changed, the old version of the program will remain available for download. 03/19/2006 - Added "trainable" neurones. Much work remains to be done, but the idea is that a trainable neurone has a sensitivity setting selected by the user, and the neurone uses this as a key for raising or lowering its threshold to meet that sensitivity in its frequency for action potentials. 02/26/2006 - Added ConnectTo and Disconnect methods to the scripted neurone's automation controller. ConnectTo allows a scripted neurone to register a line for receiving network pulses. Disconnect unregisters the line. 02/08/2006 - Finally! I/O neurones can be set to input and output across the Internet using UDP packets! Network I/O neurones allow your neural network to span multiple computers. There are no limits now! A background thread listens for incoming packets and uses the Winsock event model so CPU usage is very minimal. 10/26/2005 - Rewrote how neural scripts are handled. Before, the script was reparsed each time one of the script functions needed to be called. Since then I've learned that I only need to parse the script once, and then I can use a dispatch interface to actually call the specific function I need inside the script! Even though it was fast before, it should be quite a bit faster now! 09/04/2005 - Neural Net is available for download. ![]() |
|||
![]() |
|