I've managed to run full size Neural-Network trained on MNIST dataset of handwritten digits with 801 neurons, [(784*7 + 7*9 + 9*1) + 3] = 5563 parameters and about 81% accuracy (if i haven't messed up some calculations)...
No more no less, find out how you can also do so with your Arduino-UNO or other devices right now, by just clicking on my github link or downloading the library from arduino's "Library Manager" [...]
(just in case [about Uno] , if you encounter a verification error: first mismatch at byte 0x7800 0xff != 0x30 when uploading the sketch, then you will need to either remove an Image from the skech's NumbersInPixels or burn a smaller bootloader :P )
I think i've overtrained it, I'm an idiot whatever..