Hehe. Here is my whole plan below in clear form for all readers. Any feedback 
is helpful.

BTW Latest code, scores, text completions, and explanation of all the code can 
be found here > https://encode.su/threads/3595-Star-Engine-AI-data-compressor

New Goal:
Translation ability will allow for recognition of similar contexts so my AI can 
utilize extra seen words that came next ex. dog>meowed, and will allow advanced 
priming ability ex. cat dog pig horse ___.


Plan:
In a dataset of ex. text, similar features stick together ex. an article is on 
how to care for dogs, because humans write using the priming ability. To learn 
related words/phrases/letters, you want to look around the one of interest on 
both sides, closer is more impactful, because related words stick near each 
other! We ignore rare and too common features since they pop up everywhere or 
cost resource explosion.

Another way to find related features is an indirect context path ex. 'dog' and 
'cat' say are not seen near each other - however - you see "cat 
ate.........5GBs later.......dog ate", and so they predict the same thing, and 
therefore have a similar meaning. You can also do hole and delay matching to 
find related features ex. 'cat ate tuna on bed' and see 'dog ate kibble on bed' 
and the lengthy match proves a closer relation.

Both methods ignore rare/common features, give more impact if closer on either 
side of the feature, and need enough observations, and ex. for method2 cat/dog 
needs to share enough predictions out of the Total Contexts they have follow - 
they'll need more if there is more competing possible relations ex. if cat is 
near only 4 other words ever and you have seen cat 500 times, then you are sure 
which is closer related.


Conclusion:
I'm going to first try method2 because it seems like it would contextually find 
related features better, it uses predictions, exactly what we want to get 
precise related features. Just being near each other is powerful too, like the 
priming ability, it is the priming ability !, however this is only a nice 
enhancement and is tar dead meat without a LOT of data and compute, we will use 
method2; context predictions to be our markochaino guide mainly.

Implementation plan:
I will store in a trie tree contexts, and will take a given 2 words to check 
for similarity. I'll see how many 'word' contexts both have on the righthand 
side - they have enough experience if I see ex. cat has 5 types of words follow 
and in total has seen 500 words follow cat, combined with dog's such analysis. 
I'll take those contexts and see if they share all those contexts of total they 
have. This will tell me how similar they are and how confident it is of this 
answer. I'll ignore rare and common words by not considering them at all in the 
described steps, by seeing at the root area of the trie tree ex. 'the' has a 
count of 500 times and we have seen so far 1,000 words, "wow 'the' is really 
common in my experience" is what it will see.
------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T192296c5c5a27230-Mfe20849a30268080aef43b0e
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to