Machine learning has the potential to help us improve decision-making in high-stakes scenarios such as healthcare, hiring and social welfare. In such applications it is not enough to optimize overall performance, and instead we must provide guarantees to ensure that the systems we build have good performance across the entire population that may be affected by its predictions. At the Auton Lab, we are conducting research to characterize and mitigate fairness-related risks of algorithmic decision support. Some of the questions we are working on include: How can we identify human biases that are reproduced by a model? How can we prevent the model from reproducing and augmenting such biases? What additional forms of bias arise when the model is providing recommendations instead of making autonomous decisions?
As machine learning is applied in increasingly consequential domains, the demand for systems which can explain their decisions is increasing. From decision support systems to post-hoc, model explanations, a core research thrust at the Auton Lab focuses on interpretable AI systems. For many of our end-users, the goal of adding AI to their workflow is to augment human cognitive capacities, rather than replace humans altogether.
We develop algorithms for both detection and decision support in nuclear threat identification. Using our flagship Bayesian Aggregation method for source detection and characterization we are developing fast and efficient tools for situational awareness and safety applications. Our work focuses on robust methods, multi-sensor and multi-modal data fusion, and decision support infrastructure for rapidly processing alerts.
The Centers for Disease Control and Prevention (CDC) estimates that "each year in the United States, 1 in 6 Americans (or 48 million people) gets sick, 128,000 are hospitalized, and 3,000 die of foodborne diseases." To fight against the outbreak, and fast, it requires collaboration across agencies and a powerful data analytic platform with ultra low latency. The Auton Lab have developed such a platform that integrates the PHIS human disease database with the FSIS regulatory data to identify emerging patterns, track historical trends and predict outbreaks. The platform is successfully deployed at FSIS, used by many researchers and public health workers, and "has established a model for future interagency collaborations to maximize the value and usability of public health data".
Hospital acquired infections are a significant yet preventable detractor of patient care. According to the CDC 1 in 31 hospital patients  suffers from a hospital acquired infection. The Auton lab develops statistical models for joining disparate sources of information such as genetic tests, patient histories, geography, and other epidemiological information for detecting systematic outbreaks and identifying root cause. Leveraging multiple data sources, our algorithms establish corroborating evidence to support or dismiss hypothetical outbreak scenarios, both increasing detectability and speed of analysis while maintaining low false alert rates. References : https://www.cdc.gov/hai/data/index.html
The Auton Lab has developed several tools for intelligent fleet management and monitoring over the past 15 years. We continue to develop machine learning approaches to optimize large scale maintenance operations with a focus on reducing risks of unforeseen issues and forecasting failures.
We use machine learning tools to build various types of practical models of data. This can range from predictive models that aim to identify some interesting aspect of patient data (e.g. is a monitor alert real or artifact, are there signs of disease or not), explanatory models (e.g. what differentiates one cohort from another, or one state from another), forecasting and trending models (e.g. what is going to happen in the future, will a patient become unstable), and grouping (or clustering) entities (e.g. these patients are similar to those ones).
Counter Human Trafficking
Digital advertisements are a significant part of the modern commercial sex industry, many of which are thinly veiled ads for prostitution. Unfortunately, some of the people represented therein are victims of human trafficking. The volume and pace of this activity is beyond the ability of traditional investigative methods to consume and process. To address these challenges, the Auton Lab works directly with law enforcement and NGOs in this space to develop intelligence based tools to meet the needs of those on the front lines combating human trafficking.