I've made a ANN learn noise patterns of an animation, and playing them back, i've got a false positives problem, that i'm trying to remedy, and i actually had it going properly before, but i just ruined it.
I am planning on adding more hidden layers, but if the reconstruction from the first level is'nt perfect, then there's no point continuing further, cause it just compounds. if its a little too positive, itll go white, if its a little too negative, it;ll eventually just cull off the whole thing.
The main white image is the reconstruction as it learns, and you can see it weighs around a bit, until it then sorta hammers straight for the sequence after a while. the noise in the background is the false positives.
the synapses are visible at the start when they are bright (around the window full screen) but they all have to reduct to 0 mostly to actually make the video visible.
Edited by rouncer, 27 January 2014 - 08:47 AM.