Indian Symposium on Machine Learning (IndoML)

Title: Science-guided Machine Learning: Advances in An Emerging Paradigm Combining Scientific Knowledge with Machine Learning

Abstract: This talk will introduce science-guided machine learning, an emerging paradigm of research that aims to principally integrate the knowledge of scientific processes in machine learning frameworks to produce generalizable and physically consistent solutions even with limited training data. This talk will describe several ways in which scientific knowledge can be combined with machine learning methods using case studies of on-going research in various disciplines including hydrology, fluid dynamics, quantum science, and biology. These case studies will illustrate multiple research themes in science-guided machine learning, ranging from physics-guided design and learning of neural networks to construction of hybrid-physics-data models. The talk will conclude with a discussion of future prospects in the emerging field of science-guided machine learning that has the potential to impact several disciplines in science and engineering that have a rich wealth of scientific knowledge and some availability of data.

Bio: Anuj Karpatne is an Assistant Professor in the Department at Computer Science at Virginia Tech and his primary interest is in developing data-enabled solutions for scientific and socially relevant problems. His current research is geared towards shaping the emerging field of research in Science-guided Machine Learning (SGML) (also referred to as physics-guided ML or theory-guided data science), where machine learning methods are systematically coupled with scientific knowledge (or physics) to accelerate scientific discovery. He enjoys working on inter-disciplinary projects and his research group is actively developing novel SGML formulations in a diverse range of scientific applications including lake modeling, remote sensing, fluid dynamics, quantum science, optics, radar physics, mechano-biology, and ichthyology.

Title: Is Artificial Intelligence the new electricity that will transform Earth Systems Sciences and Engineering?

Abstract: There have been claims that AI is more profound than the discovery of fire or the invention of electricity as well as counter claims that AI will eventually destroy humanity but is not yet as mature as traditional disciplines like civil engineering. This presentation will examine the opportunities and challenges that AI may bring to bear, especially when combined with process understanding, in earth system sciences and engineering, both to advance fundamental sciences or engineering principles and for translating to societal priorities and policy imperatives.

Bio: Auroop R. Ganguly is a Professor of Civil and Environmental Engineering, as well as an Affiliate Professor of both the Khoury College of Computer Science and the School of Public Policy and Urban Affairs, at Northeastern University in Boston, MA. His research and teaching interests intersect water sustainability and climate science, urban sustainability and infrastructural resilience, as well as machine learning and nonlinear dynamics. He is a Fellow of the American Society of Civil Engineers and a Senior Member of the Institute of Electrical and Electronics Engineers. Prior to Northeastern, he was employed with the US DOE’s Oak Ridge National Laboratory and before that, at Oracle Corporation. Ganguly has published in journals such as Nature and PNAS, besides other journals across disciplines in geosciences, engineering, physics and data science, written a textbook on Critical Infrastructures Resilience, and won best paper awards in top-tier AI conferences. He is a co-founder of the Boston-based climate analytics startup risQ where he remains the chief scientific adviser. Ganguly has a PhD from the Massachusetts Institute of Technology in Cambridge, MA, and a B. Tech. (Hons.) from the Indian Institute of Technology at Kharagpur, India. 

Title: Bayesian Inference and Meta-learning for Molecular Property Prediction with Graph Neural Networks

Abstract: In this talk I will present recent work which aims to improve the generalization ability and low-data performance of graph neural networks for molecular property prediction. While graph neural networks have become a standard tool for in machine learning for molecules, thanks to their expressive power, there are two distinct challenges in this domain. The first is distributional shift: unlike in many other settings, the test molecules we would like to make predictions for nearly always differ significantly from the data used for training. Second, due to costs of data acquisition, molecular machine learning often operates in a low-data regime. As a result, graph neural networks which are underspecified by the data may not generalize well to new structure or scaffolds, and they may overfit to small training sets. We aim to address these problems by considering Bayesian estimation of graph neural networks, which captures uncertainty in model parameters and improves predictive accuracy and calibration; and by meta-learning, in which prediction tasks with few labels can benefit from other tasks where larger datasets are available.

Bio: Brooks Paige is an Associate Professor in Machine Learning in the Department of Computer Science at University College London. His research focuses on interpretable machine learning, methods for scalable Bayesian inference, and applications in the sciences. He holds a D.Phil in Engineering Science from the University of Oxford, where he was supervised by Frank Wood; an M.A. in Statistics from Columbia University; and a B.A. in Mathematics from Amherst College.