Atrix256

If you tried the demo and were remarkably unimpressed, give it a try again and it ought to work better (make sure and manually refresh - F5 so you don't have the old page cached).

I read how the MNIST data was generated and followed suit: Find the bounding box of the drawn image and scale it down to 20x20, preserving the aspect ratio. Put that 20x20 image into a 28x28 image using the center of mass of the pixels as the center point.

There is still some error due to how different mouse drawn and hand drawn numbers look, but I've found it to be a lot more accurate.

Demo: http://demofox.org/MNIST1.html

Example training data to see what i mean about difference between hand drawn and mouse drawn images: https://i2.wp.com/blog.demofox.org/wp-content/uploads/2017/0...

Thanks for the feedback on that... it was a mistake rolling it out in such a broken state.

Too funny, what's your shadertoy handle? (this is demofox)

Thanks!

BTW this older article of mine is extended with a new one that shows how to handle multiple variables (: http://blog.demofox.org/2017/02/20/multivariable-dual-number...

That's great info, thank you! (:

Interesting! I gave that a shot and definitely do hear the difference. I believe the difference you hear is due to the fact that it's a stronger low pass filter to average 3 samples instead of 2. In any case, the result is much less harsh on the ears. I'm adding that to the post, thanks!

The shadertoy was an attempt at doing the algorithm without a delay buffer (making it stateless). The real demo is the wav file and the C++ that generated the wav file (: