American Family Funding Initiative Awards

This project will prototype a machine learning framework based on data collection via mixed reality headsets and smartphones, as well as explore data processing through edge- and cloud-computing infrastructure, to generate near-real-time 3D reconstructions. The novelty of this project stems from its potential to leverage regular cameras for creating highly realistic scene reconstructions and to generate multimodal datasets by combining vision and IoT sensor devices. The systematic collection of these datasets will provide new means for property inspection and risk assessment, allowing for more efficient and accurate decision-making.

Joao Dorea (joao.dorea@wisc.edu) is an assistant professor in the UW–Madison Department of Animal and Dairy Sciences. Specializing in precision agriculture and data analytics, he applies AI to optimize farm decision making and improve animal nutrition and health.

As AI and autonomous systems are trusted with increasingly important decisions, it becomes imperative to be able to evaluate the performance of their decision-making policy before deployment. One way to accomplish this is to take data recorded while making decisions with a previously used policy and answer the counterfactual question, “What would have happened if the new policy had been making decisions instead of the older policy?” The goal of this project is to introduce novel methods for counterfactual policy evaluation in sequential decision-making settings that overcome the practical limitations of existing methods.

Josiah Hanna (jphanna@cs.wisc.edu) is an assistant professor in the UW–Madison Computer Sciences Department. His research develops and applies reinforcement learning algorithms that are effective with a limited amount of data and integrates these algorithms into complete autonomous agents. His long-term research goal is developing fully autonomous agents that learn how to achieve goals from experience.

Auto-labeling systems are emerging as a popular and widely-used alternative to reduce the cost associated with creating new training datasets. Such systems aim to reduce human labeling effort by automatically labeling data points using models that are actively trained in an iterative manner. There is very little understanding of how reliable the data obtained by such are. Confidence functions that quantify the uncertainty in a model are key components in these systems. This project will build a systematic understanding of confidence functions that can guarantee reliability for auto-labeling systems. It will also build practical algorithms based on our theoretical analysis, along with an evaluation platform to benchmark their performance.

Ramya Korlakai Vinayak (ramya@ece.wisc.edu) is an assistant professor in the Department of Electrical and Computer Engineering at UW–Madison. Her research interests span the areas of machine learning, statistical inference, and crowdsourcing. Her work focuses on addressing theoretical and practical challenges that arise when learning from societal data.

Insurance fraud costs businesses and consumers billions of dollars annually, which in turn results in premium increases. With companies relying more on automated, self-service transactions, the need to reliably detect fraudulent claims is becoming imperative. One manifestation of claims fraud is fake visual assets, such as images exaggerating damages, that have been generated or altered using AI software (deepfakes). This project will develop novel algorithms that can detect fake images from various breeds of AI generative models, even those unseen during training.

Yong Jae Lee (yongjaelee@cs.wisc.edu) is an associate professor in the UW–Madison Department of Computer Sciences. His research interests are in computer vision and machine learning. He is particularly interested in creating robust visual recognition systems that can understand visual data with minimal human supervision.

Finding highly predictive profiles from medical histories or past insurance product performance can be especially informative for improving risk assessment or evaluating new products. Sequence Analysis (SA) is particularly well-suited to handling histories of data in an interpretive way. However, a key limitation of current SA methods is that the sorting stage is unsupervised and disconnected from the predictive stage. This project will innovate on SA methods with supervised techniques that address this limitation.

Adeline Lo (aylo@wisc.edu) is an assistant professor in the UW–Madison Department of Political Science and Glenn B. & Cleone Orr Hawkins Chair. Her research interests lie in the design of statistical tools for prediction and measurement for applied social sciences. She works with and constructs tools for observational, text, experimental, and network data.

Most existing usage-based insurance uses “one size fits all” discounts and rate adjustments that fail to account for differences in policyholders’ behavior. This project will develop a framework that uses behavioral modeling to create personally tailored incentives and discounts for safer driving, using telematics data and simulation. These incentives will be calculated to reduce the expected number of future claims while adapting to the drivers’ individual levels of risk.

Yonatan Mintz (ymintz@wisc.edu) is an assistant professor in the Industrial and Systems Engineering department at the UW–Madison. His research focuses on the application of machine learning and automated decision making to human-sensitive contexts such as personalized health care.

Interpretable machine learning has become an area of significant interest due to the increased use of black-box methods such as neural networks, random forests, and others. One way of interpreting model-agnostic methods is to estimate Shapley values, which capture how much each variable (or subset of variables) contributes to prediction performance. However, one of the major challenges for Shapley value estimation is scalability, especially when the number of variables is large. This project will address this challenge by providing a reliable and scalable approach for computing Shapley values for arbitrary machine learning prediction models.

Garvesh Raskutti (raskutti@stat.wisc.edu) is an associate professor in the Statistics Dept. at UW–Madison. His research interests include statistical machine learning, optimization, graphical and network modeling, and information theory with applications to systems biology and neuroscience.

Large pretrained models must be customized for specialized applications in industry. There is no standard way to perform such customization, and existing approaches require large amounts of hand-annotated data. This project seeks to address this challenge by building a unified framework to customize large pretrained models with minimal human annotation or effort. It will develop techniques to make models compatible with any data modality, fine-tune them with minimally labeled data, and modify them to be more responsive to human feedback. This work has the potential to make powerful, general models useful for specialized industry applications.

Fred Sala (fredsala@cs.wisc.edu) is an assistant professor in the Department of Computer Sciences at UW-Madison. His research interests are in machine learning, focusing on data-efficient learning, automated machine learning, and integrating geometry into machine learning pipelines.

Electronic health records (EHRs) contain rich information regarding an individual’s medical history and health condition that can yield valuable insights to predict their life expectancy. This project aims to create a novel joint modeling approach that utilizes diverse types of EHR data, such as health event data (e.g., illnesses faced) and physiological data over time (e.g., blood pressure) from multiple individuals to accurately assess individualized mortality risk and life expectancy.

Raj Veeramani (raj.veeramani@wisc.edu) is the E-Business Chair Professor at UW–Madison, with joint appointments in the College of Engineering and the School of Business. His research focuses on new frontiers of digital transformation, industrial data analytics, IoT technologies and applications, smart and connected systems, and supply chain management.

Insurance fraud has increased dramatically with the expansion of modern technology and global communication, resulting in the loss of billions of dollars worldwide each year. Both supervised and unsupervised machine learning algorithms have been applied to detect a variety of fraudulent cases. This project will develop a novel, semi-supervised learning method for accurate and efficient fraud detection.

Jiwei Zhao (jiwei.zhao@wisc.edu) is an associate professor of Biostatistics and Medical Informatics at UW–Madison. He uses statistical methods and machine learning techniques to analyze data with massive structures.

Understanding driving behavior is central to efficient, safe transportation and associated insurance mechanisms. This project seeks to create an unbiased system for evaluating driving behavior that will use multi-modal signals, especially from audio-visual sensors, to learn contextual information about why certain behaviors happen.

Suman Banerjee (suman@cs.wisc.edu) is the David J. DeWitt Professor in the Department of Computer Sciences at UW-Madison. His primary research interests include networking and distributed systems, specifically mobile and wireless networking systems, with many applications in smart transportation, smart healthcare, and in secure and sustainable systems.

A growing number of systems, including hospitals and insurance companies, aim to derive knowledge from internal data to improve their day-to-day operations. This project will develop a causal inference framework for estimating the effects of interventions on these systems and provide algorithms to guide their use of risk prediction models in their operations, with the ultimate goal of improving services and reducing costs.

Amy Cochran (cochran4@wisc.edu) is an assistant professor at UW-Madison with a joint appointment in the Departments of Mathematics and Population Health Sciences. She works in the areas of computational psychiatry, digital mental health, and causal inference.

The most advanced sample complexity bounds on fair machine learning are called multicalibration convergence bounds. These bounds specify the number of samples required to achieve performance parity across many population demographics. This project will yield additional perspective on both algorithmic fairness and multicalibration error convergence bounds. Further, it will enable machine learning practitioners to easily understand the convergence behavior of multicalibration error for a myriad of classifier architectures.

Kassem Fawaz (kfawaz@wisc.edu) is an assistant professor in the Department of Electrical and Computer Engineering at UW-Madison. His research interests primarily include security and privacy for users interacting with their devices, with emphasis on voice interfaces, privacy policies, malware detection, data privacy, and wireless security and privacy.

One way to evaluate AI-based decision-making policies and autonomous systems before they are deployed is to take data from a previously used policy and answer the counterfactual question, “What would have happened if the new policy had been making decisions instead of the older policy?” This project will introduce novel methods for counterfactual policy evaluation in sequential decision-making, where even small changes in how decisions are made can lead to drastically different outcomes over time.

Josiah Hanna (jphanna@cs.wisc.edu) is an assistant professor in the UW-Madison Computer Sciences Department. His research develops and applies reinforcement learning algorithms that learn with small amounts of data. His long-term research goal is developing AI systems that can quickly learn new capabilities from experience.

While crowdsourcing is a popular way to collect labeled training data for machine learning, it is expensive and time-consuming to hand-label each data point. Systems that automatically label data points while actively learning a model perform well in practice, but there is no theoretical understanding of what performance guarantees can be expected from these systems, or whether the resulting biased datasets can even be trusted. This project aims to close this gap by developing theoretical foundations for characterizing the performance of auto-labeling systems.

Ramya Korlakai Vinayak (ramya@ece.wisc.edu) is an assistant professor in the Department of Electrical and Computer Engineering at UW-Madison. Her research interests span the areas of machine learning, statistical inference, and crowdsourcing. Her work focuses on addressing theoretical and practical challenges that arise when learning from societal data.

A major issue that prevents machine learning algorithms from being deployed to real-world problems is the safe handling of anomalous data that differs from the training distribution data. Prior research on out-of-distribution (OOD) data detection has been primarily driven by image recognition tasks. This project will pioneer new directions in contrastive image-language learning for OOD detection. We will explore how language and vision can provide complementary sources of information to better estimate uncertainty in claim fraud detection and other risk scenarios.

Sharon Yixuan Li (sharonli@cs.wisc.edu) is an assistant professor in the Department of Computer Sciences at UW-Madison. Her broad research interests are in deep learning and machine learning. She develops algorithms to enable reliable open-world learning, which can function safely and adaptively in the presence of evolving and unpredictable data streams.

Matrix completion, or filling in the unknown entities in a matrix, is one of the most fundamental problems in data science, and existing models are incompatible with some types of data. This project will develop a new model and algorithms specifically tailored to complete matrices with heterogeneous data, with important applications in recommender systems, computer vision systems for processing and analyzing visual images, data inference, and outlier detection.

Daniel Pimentel-Alarcon (pimentelalar@wisc.edu) is an assistant professor in the Department of Biostatistics and Medical Informatics in the UW-Madison School of Medicine and Public Health. His research focuses on robust machine learning methods to identify patterns in big and messy data. He specifically examines robust machine learning of mixtures, and linear and nonlinear structures.

Insurance products often integrate multivariate risks in design. Examples range from multi-peril policies, multi-coverage contracts, to bundled products. This project is to investigate the benefit of dependence models in product management for insurance products with multivariate risks embedded in their design.

Peng Shi is a professor in the Risk and Insurance Department, and the Charles and Laura Albright Professor in Business and Finance, at the Wisconsin School of Business. His research focuses on actuarial data science, probabilistic forecasting and predictive modeling, insurance and risk analytics, dependence models and multivariate analysis, machine learning and statistical learning, and intensive longitudinal data methods.

Machine learning and computer vision methods that drive applications such as object detection, image recognition, language understanding, and voice recognition make use of models known as “transformers” that can require weeks or months to train. Deployment of such parameter-heavy models also involves access to specialized hardware resources. This project will significantly extend the capabilities of current models based on algorithmic and implementation improvements. We will focus on ultra-long temporal/spatio-temporal sequences, coming from a broad variety of applications, and study the key challenges that need to be overcome to allow efficient training and deployment of transformer models in these settings.

Vikas Singh (vsingh@biostat.wisc.edu) is a Vilas Distinguished Achievement Professor in the Department of Biostatistics and Medical Informatics at the UW-Madison School of Medicine and Public Health. His group works on algorithm development for image analysis, computer vision, and machine learning problems motivated from applications in biomedical sciences, engineering, and industry.

Principal investigator: Colin Dewey (colin.dewey@wisc.edu), Professor of Biostatistics and Medical Informatics.
Co-Principal Investigator: Mark Craven, Biostatistics and Medical Informatics

Researchers and businesses are increasingly using large data sets, compiled from many sources, for training machine learning systems and performing statistical analyses. A major bottleneck arises from the fact that compiled data sets often contain unstandardized, unstructured metadata that describe each record. Manual standardization of metadata is labor intensive and often requires substantial expertise in the field of study.

To mitigate this issue, this project will develop machine learning approaches for automating the task of metadata standardization in large, heterogeneous data sets. The researchers will use state-of-the-art natural language processing models and develop active learning algorithms, which facilitate identification of records that would most benefit from human expert input. They will demonstrate the performance of these methods on the Sequence Read Archive—a vast repository of public biological sequence data.

Principal Investigator: Michael Ferris (ferris@cs.wisc.edu), Professor of Computer Sciences.

Uncertainty abounds in decision problems and optimization is a key tool used to mitigate its effects, utilizing the power of data science. This project will deploy a new approach that separates strategic decision making from operational modeling, in the context of a claim adjustment problem in the insurance industry. In this setting, random accidents occur across a large service area, requiring agents to deploy to the site to assess, document and determine appropriate courses of action. Our approach differentiates normal workload from crisis situations. It will inform an operational model that schedules resources over time to service both routine, normal workloads in a cost-effective manner, and enable the company to react efficiently to crisis situations. The model can be applied to problems as diverse as disaster recovery, chemical spill mitigation and electricity planning for extreme weather events.

Principal Investigator: Song Gao (song.gao@wisc.edu), Assistant Professor of Geography.
Co-Principal Investigator: Jerry Zhu, Computer Sciences.

User location information is a key component of both research and business intelligence. With the increasing availability of mobile devices and popularity of mobile apps, users in social network platforms actively share rich information about their locations on the Earth, the places they go and the activities they engage in. Those location-based profiles provide an invaluable source of information. However, mobility data is among the most sensitive data being collected by mobile apps, and users increasingly raise privacy concerns. The proposed research aims to develop a deep learning architecture that will protect users’ location privacy while keeping the capability for location-based business recommendations. The algorithms developed through this research may be applied in usage-based insurance (UBI) and other location intelligence domains.

Principal Investigator: Kangwook Lee (kangwook.lee@wisc.edu), Assistant Professor of Electrical and Computer Engineering.
Co-Principal Investigator: Dimitris Papailiopoulos, Electrical and Computer Engineering.

The recent successes of machine learning hinge on the ability of predictive models to generalize, or adapt well to previously unseen data. Data augmentation, the process of injecting artificial data points into a training set, is widely employed for improving generalization. One of the most prominent data augmentation algorithms is mixup, which helps achieve state-of-the-art generalization performance across several benchmark tasks.

While mixup algorithms are useful for improving generalization for a wide class
of tasks, they have a few critical limitations. Mixup sometimes degrades generalization, restricting the applicability of these tasks. Moreover, current mixup algorithms do not have any theoretical performance guarantees. To address these challenges, the researchers will develop a computationally efficient mixup algorithm based on a generative adversarial network (GAN). They will also develop a theoretical framework for analyzing the performance of various mixup algorithms. This research will provide a new approach to improve generalization, with provable performance guarantees.

Principal Investigator: Jeff Linderoth (linderoth@wisc.edu), Professor of Industrial and Systems Engineering.
Co-Principal Investigators: Jim Luedtke, Industrial and Systems Engineering; Daniel Pimentel-Alarcon, Biostatistics and Medical Informatics.

Matrix completion, or filling in the unknown entities in a matrix, is one of the most fundamental problems in data science. Matrix completion is used in applications such as recommender systems that predict the rating a user would give to an item, such as a movie or product, and then make recommendations to the user. This project will develop algorithms for solving a mixture matrix completion problem (MMCP), which has important applications not only in recommender systems, but also in computer vision systems for processing and analyzing visual images, data inference, and outlier detection.

Key to this research will be the development and application of advanced algorithmic techniques from integer programming, a powerful mathematical tool for solving optimization problems involving discrete choices. The work will pave the way towards the application of integer programming for a broad class of large-scale data science problems.

Principal Investigator: Michael Morgan (mcmorgan@wisc.edu), Professor of Atmospheric and Oceanic Sciences.
Co-Principal Investigator: Brett Hoover, Space Science and Engineering Center.

This project will develop an ensemble weather prediction system for American Family Insurance that will provide high-resolution weather forecasting run entirely in cloud computing infrastructure. This project will improve the accuracy of forecasting hazardous weather by producing many realizations of the same forecast from slightly varying initial conditions.

The probabilistic forecasts will provide advanced warning of not only hazards including hail, wind gusts, and hurricane impacts in targeted regions, but also the uncertainty associated with the predictability of these hazards. This novel research will provide a state-of-the-science technique in regional weather modeling.

Principal Investigator: Shivaram Venkataraman (shivaram@cs.wisc.edu), Assistant Professor of Computer Sciences.
Co-Principal Investigator: Dimitris Papailiopoulos, Electrical and Computer Engineering.

Data scientists train machine learning models that are used in a wide range of domains, from drug discovery to recommendation engines. Training a machine learning model, and fine-tuning the parameters that control how well a model performs, take significant time and resources. The process of incremental fine-tuning is often manual and involves retraining models from scratch. This project will automate and accelerate this process of fine-tuning by reusing and sharing past computations from prior training jobs, using a technique called model recycling. The researchers will develop a software framework that can help data scientists accelerate model fine-tuning, and a proposed intelligent predictor that can automatically save prior computation results, based on their importance.

Principal Investigator: Joe Austerweill (austerweil@wisc.edu), Assistant Professor of Psychology

People spend a significant proportion of their time asking each other questions to gather information. Entire professions, such as academia and customer service, are dedicated to asking and answering questions. Despite tremendous progress in machine learning, automated methods that answer a person’s questions are still inferior to answers from people.

Why are people better at answering questions? One reason is that question-askers leave out information that those answering the questions can fill in from their rich knowledge of language and the world. A recent machine learning method addresses this issue by asking multiple, reformulated versions of a human question, providing multiple answers, and learning to select the answer that is most likely to satisfy a person. However, this is done purely from data and does not incorporate psycholinguistic research demonstrating that people prefer simpler answers that are tailored to their personal goals and knowledge.

This project investigates whether incorporating psycholinguistic factors can improve automated question-answering methods. If so, then researchers can test novel, potential psycholinguistic factors and learn more about the underlying mechanisms that enable people to answer questions.

Principal Investigator: Vikas Singh (vsingh@biostat.wisc.edu), Professor of Biostatistics and Medical Informatics

Collaborators: Zhanpeng Zeng (Computer Sciences), Shailesh Acharya and Glenn Fung (American Family Insurance)

Natural language processing is a form of artificial intelligence that helps computers read and understand human language. Efficient and accurate natural language processing models are central to various applications but have a significant computational footprint.
The overarching goal of this project is to accelerate the time it takes to train and test these models by developing alternative solutions that are based on much faster image processing primitives.