I have an htm design in my head, why im interested is id like to see some feedback of its internals after its learnt patterns. Ive never seen anyone do this before, but as far as im thinking, you should be able to feedback from the final region, just rolling its prediction, but you dont pass back predictions, you just pass back spacial states.
Then you build all the intermediateness coming back out, hopefully producing some kind of movie on the eye level region, which is strange the think about what it might look like. You could roll any region, and then feedback from there, it would produce different results at each region.
Feedback I havent seen much at this point, and would be extremely interesting... even though feedback from region 0 is how you read its predictions, right?
Doesnt seem to have much to do after the predicting region rolls, you just sorta see whats contained in it.
If you start off from completely random synapse allocation, youd start off with noise as an output, but slowly what you teach it materializes inside it?
One question tho, would comparing the feedback to the internal prediction of a region come up with anything as you work your way back?
Last edit: Magnus Wootton 2013-05-17
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
There seems to be a problem with sending from layer 2 and beyond back, like the prediction must be removed before accessing the layer before... has this been thought about before?
I figure the solution would be to mark some proximal synapses as feedforward only (the predictions), and some feedback and feedforward. (the spacial pool only)
Last edit: Magnus Wootton 2013-05-17
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I have an htm design in my head, why im interested is id like to see some feedback of its internals after its learnt patterns. Ive never seen anyone do this before, but as far as im thinking, you should be able to feedback from the final region, just rolling its prediction, but you dont pass back predictions, you just pass back spacial states.
Then you build all the intermediateness coming back out, hopefully producing some kind of movie on the eye level region, which is strange the think about what it might look like. You could roll any region, and then feedback from there, it would produce different results at each region.
Feedback I havent seen much at this point, and would be extremely interesting... even though feedback from region 0 is how you read its predictions, right?
Doesnt seem to have much to do after the predicting region rolls, you just sorta see whats contained in it.
If you start off from completely random synapse allocation, youd start off with noise as an output, but slowly what you teach it materializes inside it?
One question tho, would comparing the feedback to the internal prediction of a region come up with anything as you work your way back?
Last edit: Magnus Wootton 2013-05-17
I read in another post, that you guys are actually getting motion video out of it, albiet noisily. this is really exciting! :D
There seems to be a problem with sending from layer 2 and beyond back, like the prediction must be removed before accessing the layer before... has this been thought about before?
I figure the solution would be to mark some proximal synapses as feedforward only (the predictions), and some feedback and feedforward. (the spacial pool only)
Last edit: Magnus Wootton 2013-05-17