Session C11: AI and modern statistics in HEP

Evolution of Analysis Techniques and Statistical Treatment

Experimental physicists benefit from an ever-increasing volume of data and a broad set of analysis techniques and statistical modeling approaches that have been developed over decades. With demands on analysis scale and complexity changing, so are the software tools that are being developed. I will describe how modern tools, frequently based on task graphs and employing columnar analysis techniques, affect the way in which physicists approach their analyses. I will illustrate the role of machine learning in this context with two examples. Simulation-based inference techniques take advantage of the typical HEP problem structure to allow for powerful statistical inference via machine learning. The possibility to use automatic differentiation to differentiate through an analysis pipeline enables gradient-based optimization while also introducing practical challenges.   

Evolution of Generation and Simulation Techniques in the AI/ML Era

Event generation and detector simulation are critical components of high energy physics research, with demanding computational requirements. The upcoming massive increases in data volumes and complexity from next-generation experiments, such as the High Luminosity LHC, necessitate the development of new approaches to deliver simulated events with higher efficiency. AI and ML density estimation techniques and related methods offer viable solutions for important tasks including phase space sampling, integration, and generative modeling. There is substantial complementarity between targeted and end-to-end usage of AI/ML and key opportunities to accelerate algorithm inference using coprocessors and high performance computing centers. The implications for future colliders and other beyond-next-generation experiments will be discussed.   

Evolution of Event Reconstruction Techniques in the AI/ML Era

The task of event reconstruction in particle physics data has not substantially changed in the past decades, but the solutions have. We have gone from manual tracing on photographic plates to A.I. trackers at the hardware trigger level. I will explore the evolution of reconstruction solutions in particle physics and the challenges we must address in using A.I. in the precision era of particle physics experiments.