Neural Network Extension

Question

This extension provides feedforward-backpropagation neural network functionality for Small Basic!

It allows you to easily create, train, use, save and load neural networks.

The intent for creating this was to provide a simple to use tool for teaching/learning about neural networks

Install:

Download from SkyDrive

The .zip file contains the extension's .dll, .xml plus .sb demo program code and .set training set data file.

Usage:

The extension's intellisense includes examples for every method. The demo program should elucidate how to use this extension quite thoroughly. I will add a full reference document when I have written it...

Training set data can be in text or binary format.

For text format use the following scheme:

• One value per line
• Training set items follow directly after each other
• Each item has the input values followed by the output values
• Each value gets evaluated as a double type, i.e 8 byte floating point

For binary format use the following scheme:

• Training set items follow directly after each other
• Each item has the input values followed by the output values
• Each value is an 8 byte double type

The extension has a method that allows you to convert between text and binary formats. You can use this to convert the demo training set to text format and inspect it if you're unsure of the training set data file format.

Help:

The teaching/explanation of feedforward-backpropagation neural network theory is beyond the scope of this document. You will need to hit the web and do some study if you're not familiar with the subject...

For any questions on, or problems with, the extension then please ask here and I'll see if I can help.

Answer 1

During beginning a simple HeavenOrHell (engl. Future Teller) Origami in the GW,

that should open on the horizontal or vertical line simply by a horizontal or vertical mouse gesture (eg. 2 to 3 mouse moves while observing an Xmin/Xmax and Ymin/Ymax), so that eg.:
IsVertical: 2-3 Up-Down Moves with (yMax-yMin) > (xMax-xMin) / ContinuousMove
IsHorizntl: 2-3 Lft-Right Moves with (yMax-yMin) < (xMax-xMin) / ContinuousMove
                                                                    dy          dx

, i came across this interesting extension and some usages about ANN on eg:
C# Application to Create and Recognize Mouse Gestures or

Mouse gestures recognition

Does anyone have experience with this extension and its usage for eg. a simple case like this, where only a horz/vert mouse gesture are to train and to detect.
Or maybe some code from practice that uses this extension?

Answer 2

When you look in Visual Studio samples for SB (ev. Zock77's ANN challenge suggestion in the back head),

you'll find a rather fresh Emiliano Musso article: Basis of NeuralNetworks in VB.NET with further article links, if one's interested.

The current extension from EzyMCP works well for SB 1.2 after DeReComp from C# and very few mods (some .ToString()). At least the ANN_XOR.sb demo works well.

Answer 3

(translated by Google translator)

Very interesting extension. Many thanks to the developer.

Answer 4

Would you mind posting some more examples? Perhaps a program demonstrating what this program does? NKZ435

Also, What is the sample ANN trying to accomplish?

---

It is written: "'As surely as I live,' says the Lord, 'every knee will bow before me; every tongue will acknowledge God.'" Romans 14:11

Answer 5

Can you elaborate on what training sets are?

---

It is written: "'As surely as I live,' says the Lord, 'every knee will bow before me; every tongue will acknowledge God.'" Romans 14:11