Schedule + Abstracts

10:30 am – 10:50 am: PSC tutorial (John Urbanic, PSC) | View Slides ›
10:50 am – 11:00 am: Welcome Introductions (Scott Dodelson, CMU) + Introduction to the Workshop (Rachel Mandelbaum, CMU) | View Slides ›

Session chair: Mikael Kuusela   Facilitator: Markus Rau

Title: Explorations at the Physics-AI Interface

View slides   |   View Zoom recording

Abstract: The physical sciences are replete with high-fidelity simulators: computational manifestations of causal, mechanistic models. Ironically, while these simulators provide our highest-fidelity physical models, they are not well suited for inferring properties of the model from data. I will formulate the emerging area of simulation-based inference and provide examples of how these techniques can impact astrophysics and particle physics at the Large Hadron Collider. With this in mind I will then turn to a few big picture questions: How do we incorporate our physical insight into the underlying causal mechanism into the inductive bias of machine learning architectures? Is that helpful or necessary? Why do we care if a model is interpretable? Where do we stand on the spectrum between ML-supercharged data analysis and an AI / robot scientist? How does this line of thinking influence research in AI and ML?

Session Chair: Ann Lee   Facilitator: Mike Stanley

Title: Blurring the Line Between Deep Learning and Physical Models

View slides  [Use left/right arrows keys]   |   View Zoom recording

Abstract: The upcoming generation of cosmological surveys will aim to map the Universe in great detail and on an unprecedented scale, but involves new and outstanding challenges at all levels of the scientific analysis, from pixel level data reduction to cosmological inference. As powerful as Deep Learning has proven to be in recent years, in most cases a DL approach alone proves to be insufficient to meet these challenges, and is typically plagued by issues including robustness to covariate shifts, interpretability, and proper uncertainty quantification, impeding their exploitation in a scientific analysis. In this talk, I will instead advocate for a unified approach merging the robustness and  interpretability of physical models, the proper uncertainty quantification provided by a Bayesian framework, and the inference methodologies and computational frameworks brought about by the Deep Learning revolution. In particular, we will see how deep generative models can be embedded within principled physical Bayesian modeling to solve a number of astronomical ill-posed inverse problems ranging from simple image denoising all the way to inferring the distribution of dark matter in the Universe. On the other hand, I will illustrate how, thanks to automatic  differentiation, physical simulations can be embedded as layers in Deep Learning systems, with the example of integrating a cosmological N-body simulation within a Recurrent Inference Machine, for the purpose of reconstructing the initial conditions of the Universe.

Session Chair: Rachel Mandelbaum   Facilitator: Husni Almoubayyed

Title: Deeply Uncertain: (How) Can We Make Deep Learning More Trustworthy In Scientific Measurements?

View slides  (use conference password for user AI_conf_QtC2021)

View Zoom recording  (password was distributed via Slack)

Abstract: Through the combination of large experiments and artificial intelligence (AI), we have an opportunity to imagine new avenues for accelerating scientific discovery. The complexity and size of our physics experiments — and commensurately the data they produce — are growing at a rate that outstrips the capacity of traditional computational techniques for key tasks like analysis, simulation, and decision-making. By deriving patterns directly from data, AI may provide a powerful supplement to the more traditional techniques. As demonstrated across a broad range of applications in both science and society, it is currently in its third historical wave of progress, success, and hype. While we dare to dream, there remain significant barriers to realizing the potential of AI applications for science. One of the most important issues lies at the intersection of interpretability and trustworthiness. That is, we don’t yet know how to derive a physical interpretation of error bars that are estimated with deep learning algorithms. During this presentation, we will discuss a few methods of endemic uncertainty quantification in deep learning models, including a comparison of some of the most popular methods within a simple physical context.

Session Chair: Rachel Mandelbaum   Facilitator: Andresa Campos

Title: Interactive Learning for Closed-Loop Design of Experiments, Simulations and Expert Feedback

View slides (use conference password for user AI_conf_QtC2021)    

View Zoom recording  (password was distributed via Slack)

Abstract: In most scientific domains including material physics, high energy physics and cosmology, we have control over the data collection procedure from multiple diverse sources. This inspires use of interactive machine learning algorithms that not only find input-output associations but also interact with the data generating process making intelligent decisions about what data to collect, when and how much. This talk will exemplify such interactive learning algorithms that can guide closed-loop design and integration of experiments, simulations as well as prior knowledge in the form of expert feedback. We will talk about the notions of uncertainty and model misspecification in context of interactive algorithms that use machine learning models ranging from linear, decision trees, gaussian processes to neural networks.

Session Chair: Ann Lee   Facilitator: Markus Rau

Title: Boosting Sensitivity to New Physics at the LHC With Anomaly Detection

View slides   |   View Zoom recording

Abstract: Anomaly detection techniques have been proposed as a way to mitigate the impact of model-specific assumptions when searching for new physics at the LHC. In this talk I will discuss how these techniques, when based on state-of-the-art machine learning developments, could be utilized at different stages of data processing workflow, from online to offline analysis, and the impact they could have to revolutionize the current paradigms in the search for new physics.

Session Chair: Manfred Paulini   Facilitator: Husni Almoubayyed

Title: Beyond the First Portrait of a Black Hole

View slides   |   View Zoom recording

Abstract: As imaging requirements become more demanding, we must rely on increasingly sparse and/or noisy measurements that fail to paint a complete picture. Computational imaging pipelines, which replace optics with computation, have enabled image formation in situations that are impossible for conventional optical imaging. For instance, the first black hole image, published in 2019, was only made possible through the development of computational imaging pipelines that worked alongside an Earth-sized distributed telescope. However, remaining scientific questions motivate us to improve this computational telescope to see black hole phenomena still invisible to us. This talk will discuss how we are leveraging and building upon recent advances in machine learning in order to achieve more efficient uncertainty quantification of reconstructed images as well as to develop techniques that allow us to extract the evolving structure of our own Milky Way’s black hole over the course of a night.

Session Chair: Mikael Kuusela   Facilitator: Mike Stanley

Title: Learning the Universe: Principled AI for Cosmology

View slides   (use conference password for user AI_conf_QtC2021)

View Zoom recording  (use conference password for user AI_conf_QtC2021)

Abstract: To realize the advances in cosmological knowledge we desire in the coming decade will require a new way for cosmological theory, simulation, and inference to interplay. As cosmologists we wish to learn about the origin, composition, evolution, and fate of the cosmos from all accessible sources of astronomical data, such as the cosmic microwave background, galaxy surveys, or electromagnetic and gravitational wave transients. Traditionally, the field has progressed by designing, modeling and measuring summaries of the data motivated by theory, such as 2-point correlations. This traditional approach has a number of risks and limitations: how do we know if we computed the most informative statistics? Did we omit any summaries that would have provided additional information or break parameter degeneracies? Are current approximations to the likelihood and physical modeling sufficient? I will discuss simulation-based, full-physics modeling approaches to cosmology that are powered by new ways of designing and running simulations of cosmological observables and of comparing models to data. Innovative machine-learning methodology plays an important role in making this possible. The goal is to use current and next-generation data to reconstruct the cosmological initial conditions; and constrain cosmological physics much more completely than has been feasible in the past. I will discuss the current  status and challenges of this new approach.

Session Chair: Rachel Mandelbaum   Facilitator: Andresa Campos

Title: Applications of Differentiable Programming to Fundamental Physics Research: Status and Perspectives

View slides   |   View Zoom recording

Abstract: Take the chain rule of differential calculus, model your system with continuous functions, add overparametrization and an effective way to navigate stochastically through the parameter space in search of an extremum of an utility function, and you have all it takes to find an optimal solution to even the hardest optimization problem. Deep learning, nowadays called differentiable programming, is boosting our reach to previously intractable problems. I will look at the status of applications of differentiable programming in research in particle physics and related areas, and make a few observations of where we are heading.

Session Chair: Mikael Kuusela   Facilitator: Markus Rau

Title: The Role of Statistics in Machine Learning

View slides   |   View Zoom recording

Abstract: After a brief review of the relationship between loss functions and machine learning models, I review a few approaches, including Bayesian ones, for quantifying the uncertainty in the outputs of these models and I assess the degree to which these approaches succeed.

Session Chair: Manfred Paulini   Facilitator: Andresa Campos

12:15-12:17: Welcome
12:17-12:20: Ngadiuba: Intro
12:20-12:30: Ngadiuba: Group 1
12:30-12:40: Ngadiuba: Group 2
12:40-12:42: Transition
12:42-12:45: Lanusse: Intro
12:45-12:55: Lanusse: Group 1
12:55-1:05: Lanusse: Group 2
1:05-1:15: Lanusse: Group 3
1:15-1:17: Transition
1:17-1:20: Dorigo: Intro
1:20-1:30: Dorigo: Group 1

1:45-1:47: Welcome
1:47-1:50: Nord: Intro
1:50-2:00: Nord: Group 1
2:00-2:02: Transition
2:02-2:05: Malz: Intro
2:05-2:15: Malz: Group 1
2:15-2:17: Transition
2:17-2:20: Wandelt: Intro
2:20-2:30: Wandelt: Group 1
2:30-2:32: Transition
2:32-2:35: Prosper: Intro
2:35-2:45: Prosper: Group 1