Les,

I'm comfortable with neural nets, I like the feedforward type with backprop because that's the old standby of neural nets, works great as a pattern recognizer. I trained some ChucKian little-bitty neural nets to do some music processing and they seemed to work OK,

Oh, yes, I thought your results so far looked quite interesting!

though in my implementations they might have learned the "null pattern" of becoming a pass-thru netowrk.

Could be... One of the issues is that you can never be sure exactly *what* they learned. There's a famous example of this from the old days where they set a neural net to work trying to sort pictures that had tanks in them from pictures without tanks. This worked great on the training set but when it was tested with real images it failed badly.... So they went back and it turned out that all of the training images with tanks had been taken on sunny days, those without them on cloudy ones and the net had accidentally been trained to detect sunny days.

I'd enjoy having native neural net capability so I can do more work with more neurons as implementing them in the signal chain is expensive.

Yeah... Is there a need to have them at srate anyway? That page seems to imply using them like UAnae at some triggered rate.

Maybe one day ChucK will run on multiple processors and we could dedicate some of those processors to neural net processing. Anyways, nets are fun and if you've had some intro to them then all the better.

Absolutely! Reading it back, BTW, I feel that my last mail was less then clear. I meant I'd like to approach this naively as a fun toy, not that either you or Rebecca was being naive. I hope that was clear already. Yours, Kas.