The Institute for Artificial Intelligence and Fundamental Interactions
Abstract: I'll introduce the IAIFI, The Institute of Artificial Intelligence and Fundamental Interactions, one of the NSF's inaugural AI Institutes. The IAIFI involves researchers from Tufts, Harvard, MIT, Northeastern, and Brandeis who work in several areas of physics and computer science. The IAIFI's goal is to advance physics knowledge – from the smallest building blocks of nature to the largest structures in the Universe – and galvanize AI research innovation. The efforts in this panel have come about in some way through the interactions fostered by the IAIFI. I'll discuss IAIFI's activities and offer an open invitation for others to join this cross-discipline and cross-university community
Using Machine Learning to Analyze Elusive Particle Interactions
Speaker: Jessie Micallef, IAIFI Fellow
Abstract: Neutrinos remain an elusive and intriguing fundamental particle that is useful for probing inconsistencies of the Standard Model of particle physics. Data from neutrino detectors is particularly valuable due to the neutrinos’ weakly interacting nature, thus it is crucial that we maximize the information per detected interaction. In this talk, I will show how we are using a variety of machine learning methods to better analyze the precious data from these difficult-to-detect neutrinos. I will focus on the challenges of reconstructing sparse, noisy neutrino interactions along with the advantages of using machine learning methods.
Improving unfolding using generative diffusion networks
Speaker: Hugo Beauchemin, Tufts Physics and Astronomy
Abstract: Unfolding is an inverse problem that is instantiated in many different fields, and which is solved using statistical techniques. The general objective of unfolding is to infer properties of a system, starting from data collected in an experimentation of that system. In High Energy Physics (HEP), it is used to infer the kinematic distributions of fundamental particles before they hit the detector. It allows for direct comparisons with theory predictions and is an important element of the measurement process. This presentation will cover the nature of the problem of unfolding in HEP and will illustrate it with examples taken from the ATLAS experiment. The standard technique used in ATLAS will be presented, and their limitation will be discussed. A proposal to use diffusion models to address these limitations will finally be presented.
Constrained neural networks for inverse problems
Speakers: Manos Theodosis and Alex Lin, Harvard Computer Science
Abstract: Many observable high-dimensional data are originating from complex physical systems. In many of these settings, the data generating signal is human understandable or suitable for downstream tasks, but the observable data are not. It is therefore of interest to invert this complex physical system and find a mapping from the observable high dimensional data to the (possibly low dimensional) data generating signal. We discuss unrolled optimization, a method for constraining neural networks to solve inverse problems. We show how these models arise naturally for sparse signal recovery and we extend the results to other signal models. The generality of the framework readily allows for the incorporation of equivariances into the inverse systems and we present relevant results. Finally, we demonstrate how we can incorporate uncertainty quantification into the inversion of the physical system by adopting a probabilistic (Bayesian) perspective.