BPN Neural Net in Flash - Gesture Recognition Demo

This one may be a little sensitive to url because of the way I'm grabbing the xml - so make sure you're at www.robotacid.com because Flash tends to deem that www difference as another url

Anyhow - This is Gesture Recognition for Flash. Source Code here. Just try a drawing and click on Classify to see if the neural net recognises it. You can teach it other training sets of course and find out if you can get better pattern recognition, remember to stop the training though. The advantage of Flash here is that it's easy to add graphical elements - like the error graph.

The reliability of recognition depends on the training set and the user. I'm hoping this can be applied to a game or an interactive project - but getting the training set just right is the tricky part. Plus you don't always get the same error graph on the same sample set. Just as a neural net recognises the analogue - it also requires an analogue approach to getting results out of it.

Update (07/2007): I've improved the line drawing object. The user never draws at a consistent speed so we need to force that behaviour into the data by interpolating spaced out points. I've added a line-intersection test and some comments on how to measure a straight line in FreeLine.as - don't leave it all on the neural net. I've also added random training orders - believe it or not, you only get a well trained BPN by pure luck: Some of the inhibititors on neurons may start out just right, whereas training a net with non-random elements means that a tricky sample will always be a tricky sample. Train and train again, and make sure you save that good training session - you may only get one chance. New source code is in the link above.