My commentary is pretty dopey, im sorry im a bit of a friendly dumb ass. But I finally have htm working spacially, im going to add temporal next and ill probably blow my ram budget before it starts sequencing properly I bet!
If you didnt know, htm stands for "Hierarchical Temporal Memory" and it can categorize vision bits, text bits, or audio bits, or even scalar values for statistics. It inferences, predicts and detects anomolies.
Once I get this working, im going to do a gpu version and get it as max resolution as I can, but ill keep it 1 bit depth pixels, so itll be like a hiresolution etch, which appeals to me.
The point of this movie is to show you slowly need less symantics (cells) as you group the data more, which makes the whole thing more doable on a computer.
What do I want to do with this? Id like to store a video game in the memory and actually control it on the way back out. So theres a game there, but no code driving it except the memory store!
Spark any interest?
P.S. I actually got it to feedback a little better after I recorded this by making it forget a little less harshly.
But the fact is, its clever if it could feedback like this, but it doesnt matter if it cant because its supposed to play back a stored movie, once temporal is in it, this is just accessing coincidence groups whether they are there or not, so obviously it breaks down, especially a level in.
making a hires htm reconstructor!
what would it be good for?
* labels could access its datastore. youd have to train the labels into it.
* fastforward? you could go to the nth prediction, i guess, and keep a history going back.
* it can associate its recorded behaviours to novel situations. but no new behaviours.
* it can handle interactive video.
* it can predict the future.
* it can detect anomolies in its input.
* security camera, (vitamin d video)
* statistics oracle, (grok)
but what im going to make is, * interactive video. handles action games using real footage.
I wonder if you could get close to a real dream, if you just spoon fed it enough video to actually gain some varience...