Scaling ML meeting

US/Central

Requirements on which models should be considered:

data sets don’t need to be public and we could use some toy data. We only care about scaling and the complexity of data. We have to, however, ensure that the data complexity is similar and that the models perform in a similar way with the toy data as the real data (e.g., toy data with tracking can run very quickly but takes longer for real tracks

Selected models to scale for IaaS:

Tracking as a Service (TaaS)

ITk FullSim data will be available soon

Microboone open data has been used

Tracking is cross frontier effort

IaaS can benchmark different hardware 

Memory requirements (40 GB for inference)

Tracking has multiple components: ML, CUDA kernels, etc

 

Action item:

Ask DUNE for feedback


Cosmology and imaging data is available but the models are more at the R&D stage. Will check in Cosmology expers. 

Flavor tagging: join CMS and ATLAS efforts?


CNNs in cosmology with IaaS?

 

There are minutes attached to this event. Show them.