40 MHz AD

Wednesday 5 Jul 2023, 13:00 → 14:30 Europe/Zurich

Conifer+Yggdrasil:

• Adapt data-preprocessing to allow for in-promptu selection of features for hyperparemter scan (all functions returning dataframes)
• Pass to Mathieu when finished, run full sweep on Google resources
• Figure out with Sioni and Mathieu where latency/resource estimates will enter in hyperparameter optimisation
• Look at pareto frontier real synthesis vs estimates
• Talk to Sioni about the next few wekks of work (and fast ML abstract)
• Thea ask Guenter about funding for London trip

 

DeepSets for jet tagging

• Vlad can reduce latency to 70 ns-->follow  up?
• Something is wong with the conv3, 0 ns latency--> follow up with GUI
• So latency works for 8const, otherwise resource
• Ask Javier about max partition limit and HP dataset

 

Contrastive learning

• Supervised contrastive learning: delphes SM bkg vs  SM bkg
• Encoder only w/ contrastive loss
• How to utilize in anomaly detection metric?

13:00 → 13:05 News & updates

13:05 → 13:20 Fast BDTs with TF-DF and Conifer

Speaker: Andrew Oliver

13:20 → 13:35 AXOL1TL

Speaker: Thea Aarrestad (ETH Zurich (CH))

13:35 → 13:50 DeepSets for jet tagging

Speaker: Denis-Patrick Odagiu

13:50 → 14:05 Contrastive learning and KD for anomaly detection

Speakers: Katya Govorkova (Massachusetts Inst. of Technology (US)), Mia Sodini (Massachusetts Inst. of Technology (US))

https://docs.google.com/presentation/d/1CY2ja0ISChNySDlqAPQ6fb14ENcL8m5MmfRaWnR2zXE/edit#slide=id.p