[ranch vermin]

After being thrown under the ocean with TONNES of theory and thinking,  I finally was pushed  (also by something else which will remain undisclosed.)  to go attempt an artificial vision cortex again.

If it were to be used with that tracker I made,  it would twin with it, behind it,  after the tracker has simplified the perspective data.

Ive got a whole lot more ideas under my belt,  And im pretty sure now,  I should be able to,  with AND / OR-  create some kind of primitive intellect for the finished robot to go with his motor.

Here it is->



Its showing the computer 'artificial dreaming' the frame in front, using the frames behind as a clue.   The special thing about it,  is it learns at a furious rate, and generates a lot more states than went into it,  because it is a convolution net.
This is only on cpu,  when I get on gpu ill be attempting to bring back street fighter 2 at a medium low resolution! 

[Korrelan]

I like it...  Biologically inspired too!

Looks like a cross between a convolution and a Hopefield network.

Scaling it up is going to be fun, though the pixel/ retina sensors can be spaced out over a wider area and the viewed video stream simplified... posterizing or similar filter.

 

[ranch vermin]

you mean putting spaces between the pixels?

Yes, that would give you 4 times speed up with a single gap between on the x&y,  then I guess you could glue it back together with some trickery.

Im not a biologically inspired guy,  but I did pick this stuff up off Jeff Hawkins, back in 2013,    im against animal experimentation,  so I think about it just like its some kind of video game machine with AND/OR groups.

AND = one perspective of one thing.
OR = many perspectives of the same thing.

And its how I develop collisions, of these classes/perspectives that grant its model validity,  being its intelligence, how it decides to behave on playback, due to what it saw.   And what it decided to develop itself through its methods.

[Korrelan]

So it's formulating the last frame based on what it saw prior... or generating and idea from prior knowledge.

The pixels could be replaced by any kind of knowledge/ experience, and the sequence it experiences them in would generate a prediction/ concept idea of what should come next.

Using the pixel engine to link/ vector to different types of data could be interesting. 

Edit: You wouldn't need to space the pixels in net, just the eye/ retina sensors that vector/ connect to the pixels.

[ranch vermin]

Of course.  its just when you were reconstructing it so it became visible to the "robot user" you could clean it up with a connecting filter or something.

[ivan.moony]

AND = one perspective of one thing.
OR = many perspectives of the same thing.

This is a standard technique for pattern matching when parsing texts (with modification and -> sequence, or -> choice). You can build the whole type system for analyzing data with those two.

[ranch vermin]

Thats cool you should say that.  I guess an oring could be a point of many directions,  I handle that too in this system,  but I mean orring the ANDS of sequences, before it even gets to making the choice...  and its very important they are there orring the ands, because this thing is not intelligent without them,  and its only a faint guess that its going to happen.  (basic symbolic logic, knowledge gathering.)

it stops the robot from being stuck in too much specificy - that there is many sequences of words that can arise at the following choices of segments.  so an ORRING OF ANDS TO AN OR OF CHOICES.