Coty
28-Aug-17 04:45 AM
the kind of neural networks I'm learning to design could in fact beat smb 1
atm what my stuff is able to do is simple tasks like this, as well as simple letter recognition
as in, I can write an "A" and have it guess with 96% accuracy that its an A
turns out recognizing text from images is pretty trivial, I didnt realize its basically a solved problem now
some simple math and you can design a neural net to do quite a bit, training can go very fast if you do backprop if you have test data or some other form of verifiable data
in this case I'm actually calculating the answer to the problem manually (just simple matrix math) to feed into back propagation for the network to readjust weights and biases for the synapses
but given that I have appropriate data the network could be expanded to do a lot more
my current end goal is pretty dumb atm, just seeing if I can train a neural network to play my card game, to be a better bot match for players
but getting to that point requires a deep understanding of how it all works, math-wise