Ai Dreams Forum
Member's Experiments & Projects => AI Programming => Topic started by: ranch vermin on December 15, 2014, 08:39:51 pm
-
Lets talk about an artificial vision database.
HTM is an information theory entropy and anomoly detecting system, so this at a stretch could do segmentation/classing and binding/tracking by grouping together only common results, which could be further helped with algorythms to help it know boundaries of moving things.
1. You need to be able to count instances of classes, so you know how many things are in the field of vision. Then when asking for the predicted frame, each instance will predict itself once each. (So knowing the exact count is very important.)
2. invarience in the input and output. invarience in the input being storing distances, not positions, and output storing the difference, not the positions, and then playback would be constant integration of the difference. I imagine exact positions are variables past through the synapses, to make exact frames reconstructable, But the cells fire off relative values only.
Then I can see this thing actually working, I know how motor would work, and also how to build a perceptron emulator (cpus in series) that should be able to take the punishment, easily.
It would swap between 2 modes, 1 the vision memory, and mode 2, the motor, and the motor would use 99% of the potential samples, as the vision memory is much easier to update, only once every 60hz, and motor wants 1,000,000,000 samples at least every 60hz. (Because its using evolutionary GA annealing type techniques)
-
heres an example of a translationally invarient system... im going to steal the idea and make my best video net yet.
top notch gift from this guy... (the best bit is when he tracks a head in 3d, with no depth map.... near the end.)
http://www.youtube.com/watch?v=lmG_FjG4Dy8#ws (http://www.youtube.com/watch?v=lmG_FjG4Dy8#ws)