This Week in MLC@Home
Notes for Aug 16 2020
A weekly summary of news and notes for MLC@Home

Apologies for the delay this week, We were hoping to have the new client first, but it needs another day of prep.

News:

• Again, lots of behind the scenes work this week, but next week should see a new client and new WUs
  
• MLDS v9.50 should be released in a day or two, including official ARM32/64 support, as well as other fixes
  
• Thanks to those who were able to test the beta ARM clients!
  
• New Dataset 3 WUs should also roll out this week, but will require the new client
  
• Shared the results so far with other researchers, and received positive feedback on the direction and insights provided. Paper on the results so far is under way
  
• No updates on the new server.
  

Project status snapshot:

Tasks
Tasks ready to send 38906
Tasks in progress 17358
Users
With credit 499
Registered in past 24 hours 19
Hosts
With recent credit 1565
Registered in past 24 hours 21
Current GigaFLOPS 23110.73

Dataset 1 and 2 progress:

SingleDirectMachine       9680/10004
EightBitMachine           9634/10006
SingleInvertMachine       9684/10003
SimpleXORMachine          9647/10002
ParityMachine              358/10005
ParityModified              44/10005
EightBitModified          1549/10006
SimpleXORModified         6480/10005
SingleDirectModified      6460/10004
SingleInvertModified      6530/10002 

Thanks again to all our volunteers!

-- The MLC@Home Admins

Thrilled to read about the upcoming dataset 3 and thus the next round of models we are going to train. Any primer of what aspects you will focus on in this round? How will this subsequent training round differ from previous ones?

This round will focus on a wide net. Dataset 3 will have 100 difference random finite automata, and train multiple identical networks on each of these networks, and see if we can differentiate between each network. This is a much harder problem than differentiating between our current set of 5.

The next client update (not the one rolled out tonight, but the next one) will introduce support for simple CNN based networks as well RNNs, so we can do things like train on CIFAR10 or MNIST-based image datasets. I'm very interested in using the generation tools from TrojAI to see if we can differentiate between networks trained with their compromised versus pure training data.

I know you in particular were interested in hyperparameter search, but that's not what we're focused on at the moment. When school starts in the fall, I'm hoping to bring more researchers on board to maybe tackle that problem.

Thanks for getting back to me! Appreciate that very much!

very interesting to know what's coming next to MLC@Home and increasing the complexity of the trained networks might yield more compelling evidence for your initial NN-differentiation hypothesis based on its training set. Will we also likely see an increase in runtime of the future WU-set?

Exciting to see support for CNN and RNN, as they are definitely more prevalent NN structures nowadays. As I really don't know much about the programming and porting of the applications, I am really in awe how fast you keep progressing with the project here. TrojAI sounds like an awesome tool and very much tailored to your research purpose. Amazing what's already out there :)

Hyperparameter search and optimization might also be more difficult to implement, thus I fully understand your decision not only from the perspective of your current research interest. Reading between the lines however, I see that it still might a valid research area down the road, so that's exciting!