On Mon, Jun 7, 2021, 1:47 PM <[email protected]> wrote:

>
> https://encode.su/threads/3635-Hutter-Prize-Entry-quot-STARLIT-quot-Open-For-Comments
>

I am testing it now. It should win the Hutter prize if it passes, but won't
claim the top spot on the large text benchmark because of the CPU time and
memory constraints.

The trick is to reorder the articles in enwik9 to maximize mutual
information between adjacent articles. This matters more in constrained
memory contests than when the complete model can fit in RAM. Then the text
is preprocessed using the dictionary from phda9 and compressed using a
modified version of cmix with parts removed to meet the requirements of
using 10 GB memory and 50000/(geekbench 5 score) hours. For my 2.8 GHz
Lenovo i7-1165G7 with a geekbench 5 score of 1427, that's 70 hours in one
thread. I expect it to take 50 hours based on smaller test files. The
original cmix takes a week with 32 GB.

cmix uses a LSTM neural network. The leader, NNCP, uses a transformer
network on a GPU, which is not eligible for the Hutter prize.


------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T950e4979da3bfbbf-M540c46fbc5ccc9888122b6f6
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to