[ranch vermin]

Its a binary tree where every step you flip a coin, but you bias one leaf over the other...  its actually a binary markov chain.

I think this format,  you can store the activity of any program inside it.  (especially animation.) 

If you build up a symbolic vocabulary,  you can store it temporally inside one of these things (exactly like a markov chain)
And you should get the probability of any state that it saw.   The length of the chain is the attention span,  and you train it
once per frame,  inserting the state, then every time you train it (the whole length of its memory) you increment the directions you
went till the end.

When playing back, you insert the data you have, and then find the next state by flipping a coin, then biasing the result with the connection weight,
if you get to the point where there is no state.  you skip a letter in front, and then grab an available probability, if its not available, you skip ahead again,
you make time back, by going as deep as you can all at once, and do less whole calls.

It shows the same lack of sense as a markov chain chat bot. (Because its the exactly same theory, except its binary.)   But if you tokenize something, you
can then produce back what you saw statistically.  Itll need spacial invarience, and order + dynamicy invarient,  if you want to insert into it different amounts of things, and still get use out of all your chains,   you may have to teach it the same thing in many orders, so it becomes resistant to useless order.


One network should be able to contain every bit of the program you are watching and tokenizing,  if its object oriented, you can play back and train each object in parallel.


Id love to see what this thing could do with audio samples. (you go to the harmonic domain to simplify the data.)   Then its a pure audio stream Markov chain.

[ranch vermin]

And heres an example of what its like ->  (i know its old, but its the best thing!!!)

store it temporally inside one of these things (exactly like a markov chain)
And you saw statistically.  Itll need spacial invarient,  if you want to insert into it different amounts of things, and still get use out of all your chains,   you can store the activity of any state by flipping a coin, then every bit of the chain chat bot. (Because its the exactly same thing in many orders, so it becomes resistant to useless order + dynamicy invarience, and order + dynamicy invarient,  if you want to useless order.

It shows the exactly like a markov chain chat bot. (Because its the exactly same theory, except its binary.)   But if you get the program you are watching and tokenizing,  if you want to useless order.

[ranch vermin]

I just learnt more about it,   and now i realize i dont know didly squat.   that you can actually not just copy the sensor, but derive an NLP on the frontside too if you want.   sheevers.