Choose timezone
Your profile timezone:
HEP Journal Club Seminar
Fast Machine Learning Inference for Scientific Discovery
by
→
Pacific/Honolulu
Room 417A (Watanabe Hall)
Room 417A
Watanabe Hall
20
Description
The very high raw data flow rate at the Large Hadron Collider (LHC) poses a significant challenge to search for physics beyond the standard model at the ATLAS experiment. Currently, we use a two-stage data filtering in the trigger system comprising the initial low-level hardware-based trigger and a subsequent software-based high-level trigger. The big data challenge is anticipated to be even worse in the future High-Luminosity phase of the LHC. In this talk, I will highlight the potential applications of ML for hardware (ASIC or FPGA) triggers to mitigate some of these challenges. I will discuss one strategy to implement a neural network model on Field Programmable Gate Arrays (FPGAs) using the hls4ml software package. hls4ml is a user-friendly software based on High-Level Synthesis (HLS) designed to deploy neural network architectures on FPGAs. I will discuss one of my recent works on Transformer-based algorithms and its potential application in the trigger. But, the implications extend far beyond particle physics. I will elucidate the broader significance of low-latency, high-throughput computing challenges across various scientific domains and demonstrate how fast ML inference can mitigate such challenges to accelerate scientific discovery.
