Deep Neural networks (DNN) are known to be vulnerable to adversarial inputs. We describe some recent results towards building more robust DNN classifiers in two settings: (1) physical perturbation attacks on real-world objects and (2) deepfake detection. On (1), we describe a system called GRAPHITE that automatically and efficiently generates candidates for robust physical perturbation attacks on hard-label blackbox classifiers. Our hope is that GRAPHITE can help lead to advances in defense against robust physical perturbation attacks, which remains an open problem. On (2), we address a significant challenge that adversarial perturbations can be misused by deepfake designers to overcome state-of-the-art deepfake detectors.

In this talk, I will briefly summarize a few data driven approaches that we have developed for space weather forecasting, including techniques for solar flare predictions and quantified terrestrial impacts of major space weather events.

Artificial Intelligence has been applied in several ways to make progress in understanding planets around other stars, placing our Solar system, and the potential for life that it represents, in context within our Milky Way galaxy: a) advanced image processing to remove light from the central star to find planets; b) efficient modeling of planet spectra in high dimension spaces; and c) discerning subtle correlations between planet properties in several dimensions in order to build predictive theories of formation and evolution.

Animal movements are key to understanding many biological phenomena. The largest movement dataset from primates derives from 25 collared baboons in Kenya across a two-week timeframe (a dataset that single-handedly produced two Science papers and a handful of other publications). However, deploying on-animal-sensors in primates is extremely difficult and even where it can be done, it may not be ethical if it puts animals at risk. We envision a landscape that uses AI to do the heavy-lifting for us. We propose to outfit a forest (not the animals) with the ability to monitor the movement of individual primates. We have a small, tractable forest in Costa Rica where we can track animals using a grid of acoustic sensors placed throughout the landscape. Because primate vocalizations have individual signatures, deep learning can be harnessed to identify individual monkeys as they traverse across the landscape. The potential for understanding the social dynamics within and across animal groups in their natural habitats would be unparalleled.

Complex and computationally expensive simulations are increasingly used to probe complex systems that are inaccessible through experimental approaches. However, these simulation tools are rife with uncertainty: unknown parameters, unknown initial conditions, model errors -- these all contribute to significant risk for using simulations for important decisions. In this lighting talk we discuss emerging techniques for computationally enabling uncertainty quantification at scale.

Rapid advances in artificial intelligence (AI) have the potential to significantly increase the productivity, quality and profitability in manufacturing systems. Traditional mass-production will give way to personalized production, with each item made to order, at the low cost and high-quality consumers expect. Manufacturing systems will be resilient to multiple disruptions, from small-scale machine breakdowns, to large-scale natural disasters. Products will be made with higher precision and lower variability. Gains have been made towards this vision of Industry 5.0, a sustainable, resilient, and human-centric manufacturing system that uses AI as a tool. Despite early successes, challenges remain to realize this vision.

Molecular properties depend not only on chemical connectivity (bonds expressible in a graph), but also on the broad geometric space involving angles and torsions between these bonds. Sampling of this space is considered a grand challenge for computation due to the combinatorial expansion in number of conformers with size of molecule. In this talk, we present a modern reinforcement learning approach to the conformer sampling problem – which is trained via a carefully designed curriculum –and discuss the principles behind this strategy.

How can we detect an asteroid the size of Ann Arbor at twice Neptune's distance from the Sun---an asteroid we could potentially visit with a spacecraft? I'll describe the DECam Ecliptic Exploration Project (DEEP), a UM-led astronomical survey intended to discover thousands of the faintest solar system objects ever detected from earth. We have developed a novel approach to moving-object detection in astronomical images that uses machine learning to reduce backgrounds by roughly a factor of one million. I'll describe how this technique can be extended to even fainter objects by combining data from multiple nights and even multiple telescopes. In this way, we hope to discover a flyby target for NASA's New Horizons spacecraft, which flew by the Pluto system in 2015 and is now passing through the outer regions of the Kuiper Belt.

Advances in numerical algorithms and computing power bring digital twins to the forefront of operational analysis of many systems. Typically, digital twins are developed based on physics-based first principles and require various parameters to be specified. Some of these parameters are not observable in physical systems. When physical laws to identify those parameters are unavailable, educated guesses are employed, but inappropriate assumptions can cause substantial deviations in a digital twin's outputs from the actual system. To make digital twins represent near-exact replicas of real systems, we leverage the power of Big Data to identify those unknown parameters with observational data.

The computational power at the edge device is steadily increasing. AI chips are rapidly infiltrating the market. Tesla's autopilot system has compute power equivalent to hundreds of Macbook pros, and small local computers such as Raspberry Pis have become commonplace in manufacturing. This change opens a new paradigm of AI-driven data analytics for connected systems within IoT for cross-learning and optimal decisions. We will discuss our efforts in data analytics aimed at bringing this future into reality with applications in manufacturing and sustainable energy systems. AI techniques for transfer learning and federated uncertainty quantification and feature extraction will be highlighted.