I've implemented first machine which recognizes digits. I have used
U.S. Post Database of digits to train the machine, and I have also used
their database to test the machine. For each row, first number represents the digit and the next 256 numbers represent the input matrix (16x16).
First impression is that ActionScript is too weak for numeric computation. In the implemented machine I've got 256 input neurons, 20 hidden neurons on a single layer and of course 10 output neurons. Of course I've tried to increase the number of neurons on hidden layer but this seemed to be unreliable to actionscript 'cause AIR application just blocked after few seconds of computation. All after all, with the previous configuration I have succeded to train the machine to recognize ~1800 of 2007 digits from tests (the best case was 1815, but in general, over 1800) after a few trainings (generally after the second train). See the example
here. It would take more than 10 seconds to train the machine but test would run great (works on my machine).
I must admit that the computation in getting results (even in many tests like those from this file) is done fast enough in ActionScript so I am sure that a component developed in Flex would be capable to run in normal conditions if it had the weight matrix computed previously. I am afraid that a learning machine would be ineffcient if it has been implemented in ActionScript so this is the reason why I think matrix of weight computation in backpropagation algorithm should be done in another layer using another programming language (e.g. Java and OpenAMF).
For those who are sceptic, an example of application for digit recognition: a replacement for captcha texts (draw a digit and we're sure you're not a robot).
Here you have the source code for NeuralNetwork class. Suddenly, comments are lost when the file was uploaded on ftp :D (coming up)