LPC Coffee Hour

Thursday 13 Oct 2022, 13:00 → 14:00 America/Chicago

Sunrise (WH11NE) (Fermilab)

Aidan Grummer (Fermi National Accelerator Lab. (US)), Alexx Perloff (University of Colorado Boulder (US)), Kaori Maeshima (Fermi National Accelerator Lab. (US))

What is it? An informal chat with (healthy) food and coffee/tea to discuss anything interesting at the LPC.

When can I start eating? Coffee, tea, and snack service will start 15 minutes before the beginning of the event.

What you need to do? Audience: Prepared to be social. Presenters: No slides please, but you can use the white/smart board or bring cue cards.

Where can I find other Coffee Hour events? Check out our website!

For information about how to get on-site, please got to the Campus Access & Experience webpage.

13:00 → 14:00 Ultrafast Neural Networks for on-detector data processing

AI Machine learning (ML) has more recently been demonstrated broadly in physics, among many other domains, as a powerful tool for data compression, processing and analysis. Highly efficient AI instrumentation relies on co-design: the idea that system constraints, algorithm development, and hardware implementation inform and guide each other in complementary ways. In addition to real-time latency constraints, other experimental challenges include operation in extreme environments with limited power and area budgets. The development of on-chip efficient AI implementation begins with algorithm performance and design, which must consider the size and power constraints of the system. ML uses large data sets to create a model which excels in tasks requiring pattern recognition, anomaly detection and prediction. Neural networks (NN) are a class of models within ML, which are generally characterized through a number of multiplication and addition operations using fitted parameters (weights) determined during the training phase. By reducing both the number of mathematical operations and how often the weights need to be accessed, the algorithm can be made more efficient. Furthermore, the precision at which the calculations are performed is also important and determines the computation accuracy. Our discussion will be focused on developing tools, methodologies and demonstrators for highly efficient on-chip implementations of real-time machine learning algorithms to enable data processing at source, while operating in extreme radiation and resource constrained environments.

Speakers: Farah Fahim (Fermilab), Farah Fahim (Fermi National Accelerator Lab. (US)), Jim Hirschauer (Fermi National Accelerator Lab. (US)), Nhan Tran (Fermi National Accelerator Lab. (US))