Continual Learning for High-Energy Physics– Marco Barbone (Imperial College London)

Thursday 12 Oct 2023, 12:30 → 13:30 America/Los_Angeles

48/2-224 - Madrone (SLAC)

In many real-world applications, neural networks (NN) are trained offline on large datasets and then deployed on specialized hardware for inference. This traditional approach separates the training and inference phases. However, in practical scenarios, the training environment may differ from the real-world setting, or data may arrive in a continuous stream, constantly changing. In such situations, there's a growing need for the ability to continuously train and update NN models.

 

Continual Learning (CL) algorithms are designed to address this need by enabling the training of models on a continuous stream of data. This is particularly relevant in high-energy physics experiments where intelligent detectors are being developed. These experiments involve running algorithms on systems close to the detector to handle the challenges posed by high data rates and changing conditions.

 

One significant challenge in this context is that detector conditions can change over time, potentially causing NNs to lose accuracy or even miss important events. CL offers a solution by utilizing large, continuously streaming data to enable the network to recognize changes and adapt to evolving detector conditions. This adaptability is crucial because not all possible scenarios can be predicted and modeled in static training data, making CL a promising approach to enhance the performance of NNs in dynamic, real-world environments.

Join from PC, Mac, Linux, iOS or Android: https://stanford.zoom.us/j/98973156241?pwd=cEU5RFdlVXoyc0JTeTlDMkozKzQ5UT09

2023-10-12 SLAC seminar.pdf

video1074672244.mp4