Its only spacial feedback, but I think im actually ready to add temporal now its not conflicting, just I have a little problem with temporary flashons that unlearn after they learn, I bet! you can add motor to this thing and control the feedback with your joystick! :)
haha fun fun.
I guess all Ive achieved so far is separating my sprites from each other over the dendrites successfully.
But its an achievement already, I definitely want to see what recording the most common coincidences does, its mystifying to me.
Last edit: Magnus Wootton 2013-06-01
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
One problem I have though, is what happens when you run out of valid columns for more different spacial patterns? Ive got a little mario brothers game recording into it now, its working, but how long for before you just run out of dendrites?
anyone else thought about this problem?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Adding temporal got it up to about 1000 lines.
but its only predicting a cell once a blue moon, so somethings wrong, and you run out of cells in the column in 1 seconds... i tried boosting cells in the column to 333... my idea is tho, you can reduce permutation space and you can maybe fully occupy a 4x4 pixel area instead? then put these in a hierarchy, and maybe you get more feesable results and more predictions.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
utube link -> http://www.youtube.com/watch?v=RtMou3RM01s
Its only spacial feedback, but I think im actually ready to add temporal now its not conflicting, just I have a little problem with temporary flashons that unlearn after they learn, I bet! you can add motor to this thing and control the feedback with your joystick! :)
haha fun fun.
I guess all Ive achieved so far is separating my sprites from each other over the dendrites successfully.
But its an achievement already, I definitely want to see what recording the most common coincidences does, its mystifying to me.
Last edit: Magnus Wootton 2013-06-01
One problem I have though, is what happens when you run out of valid columns for more different spacial patterns? Ive got a little mario brothers game recording into it now, its working, but how long for before you just run out of dendrites?
anyone else thought about this problem?
http://www.youtube.com/watch?v=e6-dgLSqzIA&feature=youtu.be well its not really working... heres the result.
http://www.youtube.com/watch?v=ub7NITAlZsA&feature=youtu.be
got feedback working! :D
Interesting experiment with 2 levels!
http://www.youtube.com/watch?v=4mIyj6ZKgsM&feature=youtu.be
Adding temporal got it up to about 1000 lines.
but its only predicting a cell once a blue moon, so somethings wrong, and you run out of cells in the column in 1 seconds... i tried boosting cells in the column to 333... my idea is tho, you can reduce permutation space and you can maybe fully occupy a 4x4 pixel area instead? then put these in a hierarchy, and maybe you get more feesable results and more predictions.
I just got an idea, if you keep overwriting the "least used" cell, maybe thats the most appropriate way to use your multi cell per column!
ill come back with results if I get anywhere with it!
Last edit: Magnus Wootton 2013-06-09