Oops, I had some new abilities ON, which means the no preprocessor mode had a 
worse score (because it could not use the new abelites without changing 
weights, time taking lol), making it look like a shuffled preprocessor alone 
got it done a lot. New tests:

*Let me compare to above's, maybe no change! So we see 0.5"MB" increase when go 
up to the shuffled run for the smaller input, and 0.37"MB" for 1MB input. For 
the below, it is 0.38"MB" and 0.27"MB". The amount shaved off for just going 
down the shuffled run for both above and below post is 3.5"MB" and 1.9"MB" for 
above post and 3.1"MB" and 1.6"MB"for below. So, by just looking at the below 
visually, it looks, good still! Woho.*

without preprocessor
28,081
240,741

with [shuffled] preprocessor
25,018
224,552

with preprocessor
24,641
221,854




@Matt

"Dictionary preprocessing helps by reducing the input size, which reduces 
memory usage,"

"and by effectively increasing the context length without increasing the number 
of contexts that need to be mixed."

I'm going to reverse engineer it to be online (on the fly).

More than 16 letter exact-matches don't help much, little own as much as above, 
and this is because it is rare to find many or even one 20 letter context 
match. So it isn't increasing the context size for me (I just ran tests above 
with no hole or delay matching or aheadoftime predictions, only exact matching, 
priming, exponential functions, meaning I did not do longer than ex. 20 letter 
matching by using hole matching etc).

There's only 2 possible explanations for why cmix preprocessor is giving me so 
much shaved off score still:
1) Because it is predicting multiple letters when it predicts a letter, similar 
to Byte Pair Encoding style.
2) Because it is paying attention to context matches at the spaces and common 
joints, to get predicted next letters, ex. it asks what is the next letter for 
we walked, walked, ed, and no context match. Instead of walking alking lking 
king ing ng g and nocontextmatch.

I'm going to be testing those 2 to see which or if both are helping and how to 
improve them.
------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T192296c5c5a27230-M8f52a461e2f7528aea290490
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to