This Week in MLC@Home
Notes for Dec 28 2020
A weekly summary of news and notes for MLC@Home

Summary
Happy holidays and happy new year from MLC!

So we decided to put out an update this week anyway,, First, we've completed milestone 2 of Dataset 3, that means 1000 trained networks for each of the 100 randomly generated automata! This is a huge milestone and makes analysis much more interesting! Another important user-visible change is the website and project now sit behind a CDN. This vastly reduces the bandwidth requirements on the server, and speeds up client downloads tremendously, and there appears to be few negative effects on the project as a whole. If you see issues, please post them in the forum so they can be addressed.

Finally, we've noticed a large influx of new volunteers over the past week. Welcome! And thank you for supporting the scientific work we're trying to do.

More news below:
* We are working on revamping much of the behind-the-scenes machinery for validation and assimilation to update our handling of NaN.
* Our updated WUs to try and cross the 1000-network threshold for all DS2 WUs seems to be having a positive effect, with nearly 100 new networks complete within the past week. We will continue this until we cross that threshold.
* Over the break, our main effort is to continue to work on analysis, paper writing, a new ROCM client release, and continue to prepare for DS4.

Project status snapshot:
(note these numbers are approximations)






Last week's TWIM Notes: Dec 22 2020

Thanks again to all our volunteers!

-- The MLC@Home Admins(s)
Homepage: https://www.mlcathome.org/
Twitter: @MLCHome2

It's days that i'm crunching only "rand_automata" wus. Is it normal??

If you're doing CPU, yes. Congrats, you helped push us over the DS3/milestone 2 threshold, and are now working on milestone 3!

DS1 and DS2 are nearing completion (we may stop the Parity* WUs from each one once we get 1000 instead of 10,000, since they take so much longer to compute, and 1000 may be enough for the analysis), so as we stated last week and this week, we're flooding the remaining DS1/DS2 tasks on the GPU queue, upping the number of epochs on the GPU queue alone (so the WUs take longer to complete), and freeing CPUs to crunch on DS3 tasks. We'll revisit once DS2 goes above 1000k complete. This is mainly driven by wanting at least 1000 examples of each network type for the paper, and DS2 ParityModified is the only one not at that threshold yet.

We've been doing this for 2 weeks now, and have a) completed all the remaining EightBit WUs in DS2, b) move ParityMachine over 1000 trained networks in DS1, and c) completed over 120 ParityModified DS2 WUs in that time, compared to just ~400 total in the past 4 months. So the strategy is working out for the science. Plus, DS3 WUs generally use less resources (more compute, less disk and RAM) and are better suited for running 4/8/12/16/24/32 in parallel like we have on a lot of hosts (vs 1 or 2 GPUs).

DS4 WUs will go to both GPU and CPU again.

Especially excited about the prospective paper and the launch of the new experiment (ds4) in the near future. Congrats on gaining traction and growing the voluntueer base.

I do really appreciate that you further look into NaN-error handling as this seems to be getting more of an issue by the day.

Happy new year to all of you as well! May 2021 become a better year for everyone! Cheers!