Hridesh Rajan, Professor and Chair of Computer Science, Iowa State University
“Dependable Data Driven Discovery Framework”

Data-driven decisions are permeating nearly all aspects of our society and daily lives, but can we depend on these decisions? This talk will describe the D^4 (Dependable Data Driven Discovery) framework, a structure to organize, formalize, and understand the dependability of data science lifecycles. In particular, we will discuss the overarching dependable data driven discovery framework: risks, measures and mechanisms and show some case studies for applying the framework.

Dr. Zhengwei Hu, Data Scientist, System Health Analytics Group, John Deere
“Application of Machine Learning to Reliability Analysis in Smart Industry”

With the emergence of advanced machine learning algorithms, smart systems and industrial intelligence have dramatically changed the traditional way of production, which is replaced with real-time monitoring, modeling and prediction based on big data, cloud computing and high-performance machine learning models. This innovative technology not only enables engineers to design high-reliability products and deliver better predictive maintenance, thereby greatly shortening the cycles for component replacement due to precise failure prediction, low latency expert alerts before failure occurs, but also enables stakeholders to reduce cost, optimize product strategies, and make smarter decisions. Two topics will be covered in this talk. Firstly, a system reliability prediction model integrating physics-based and machine learning-based method will be discussed, which explains how to combine traditional First-Order Reliability Method (FORM) with Support Vector Machine (SVM) to accurately predict potential system failure. Secondly, a multi-class classification model using Recurrent Neural Network (RNN) and Natural Language Processing (NLP) will be introduced, which is used to recognize the patterns of machines and to predict the most likely failure type for each machine.

Bijaya Adhikari, Assistant Professor of Computer Science, University of Iowa
“Data Driven Real-Time Epidemic Forecasting”

The currently unfolding COVID-19 pandemic has highlighted the necessity of robust real time epidemic forecasting models. It also has led to a maturing of methods for epidemic modeling and forecasting with the CDC establishing the first Center for Forecasting and Outbreak Analytics in 2021. A variety of forecasting innovations in machine learning and deep learning were developed, with many lessons learned for COVID-19 and future pandemics. In this talk, I will highlight two specific data driven epidemic forecasting models, CALINET and DEEPCOVID, which were designed to predict influenza cases in presence of COVID-19 and COVID-19 incidence, respectively. These models have regularly performed well in real time forecasting (top 5 in CDC COVID-19 Forecast Hub) and have won multiple awards.

Aleksandar Poleksic, Professor of Computer Science, University of Northern Iowa
“On the Geometry of a Biological System”

Modern machine learning methods, such as deep learning and matrix factorization, have helped gain insight into disease biology and drug mechanism of action. These insights, however, can be attributed more to the development and applications of advanced mathematical and statistical techniques and less to understanding of the true nature of the biological space. By representing biological entities as points in a low dimensional Euclidean space, machine learning methods are simply assuming the flat geometry of the biological space. On the other hand, several recent theoretical studies suggest that biological networks exhibit tree-like topology with a high degree of clustering. If this is true, embedding a biological network in a flat space will inevitably lead to distortion of the distances between biological objects. To test these claims, we have developed a novel methodology for drug-target interaction prediction that uses the hyperbolic space as the latent biological space. When tested against its Euclidean counterpart, our method exhibits superior accuracy while lowering the embedding dimension by almost an order of magnitude. We see this as concrete evidence that the hyperbolic space can accommodate exponential growth in the number of network features while properly modeling the proximity between biological objects. Hence, our study provides additional proof that the native biological space is not the Euclidean space but the hyperbolic space of negative Gaussian curvature.