BEGIN:VCALENDAR
VERSION:2.0
PRODID:icalendar-ruby
CALSCALE:GREGORIAN
X-WR-TIMEZONE;VALUE=TEXT:Europe/Zurich
BEGIN:VEVENT
DTSTAMP:20261014T204707Z
UID:380c7c4f-4a70-40f2-833e-ea3cfd50cc84
DTSTART:20261027T130000
DTEND:20261027T141500
DESCRIPTION:Experimental data from the Large Hadron Collider (LHC) poses an
  extreme challenge for machine learning: transforming high-dimensional sen
 sor data into meaningful physical representations at unprecedented through
 put and latency. The CMS experiment at the LHC processes data streams equi
 valent to several percent of total global Internet traffic\, requiring neu
 ral network inference within tens of nanoseconds to identify rare physics 
 processes among trillions of collisions.\n\nIn this talk\, I will discuss 
 how we develop machine learning methods that operate under these constrain
 ts. We combine hardware-aware neural network design\, quantization\, and s
 pecialized computing architectures to enable ultra-low-latency inference o
 n FPGAs and ASICs. In parallel\, we use physics-informed representation le
 arning to extract compact\, expressive features from high-dimensional dete
 ctor data\, preserving information relevant for downstream inference and s
 ensitivity to previously unseen phenomena. I will highlight how learned re
 presentations\, efficient inference\, and anomaly detection can be integra
 ted into real-time processing pipelines\, enabling increasingly model-inde
 pendent searches for new physics at the LHC and future experiments.
LOCATION:ETH Zürich\, Hönggerberg HIT Siemens Auditorium (E 51)
ORGANIZER:Thea Klaeboe Aarrestad
SUMMARY:Learning under extreme constraints: Nanosecond inference and physic
 s-informed representations for physics discovery at the LHC
URL;VALUE=URI:https://thaarres.github.io/
END:VEVENT
END:VCALENDAR