Developers and academics have grown increasingly interested in developing “trustworthy” artificial intelligence (AI). However, this aim is difficult to achieve in practice, especially given trust and trustworthiness are complex, multifaceted concepts that cannot be completely guaranteed nor built entirely into an AI system. We have drawn on the breadth of trust-related literature across multiple disciplines and fields to synthesize knowledge pertaining to interpersonal trust, trust in automation, and risk and trust. Based on this review we have (re)conceptualized trustworthiness in practice as being both (a) perceptual, meaning that a user assesses whether, when, and to what extent AI model output is trustworthy, even if it has been developed in adherence to AI trustworthiness standards, and (b) context-dependent, meaning that a user's perceived trustworthiness and use of an AI model can vary based on the specifics of their situation (e.g., time-pressures for decision-making, high-stakes decisions). We provide our reconceptualization to nuance how trustworthiness is thought about, studied, and evaluated by the AI community in ways that are more aligned with past theoretical research.

Hail causes billions of dollars in losses by damaging buildings, vehicles, and crops. Improving the spatial and temporal accuracy of hail forecasts would allow people to mitigate hail damage. We have developed an approach to forecasting hail that identifies potential hail storms in storm-scale numerical weather prediction models and matches them with observed hailstorms. Machine learning models, including random forests, gradient boosting trees, and linear regression, are used to predict the expected hail size from each forecast storm. The individual hail size forecasts are merged with a spatial neighborhood ensemble probability technique to produce a consensus probability of hail at least 25.4 mm in diameter. The system was evaluated during the 2014 National Oceanic and Atmospheric Administration Hazardous Weather Testbed Experimental Forecast Program and compared with a physics-based hail size model. The machine-learning-based technique shows advantages in producing smaller size errors and more reliable probability forecasts. The machine learning approaches correctly predicted the location and extent of a significant hail event in eastern Nebraska and a marginal severe hail event in Colorado.

Severe weather, including tornadoes, thunderstorms, wind, and hail annually cause significant loss of life and property. We are developing spatiotemporal machine learning techniques that will enable meteorologists to improve the prediction of these events by improving their understanding of the fundamental causes of the phenomena and by building skillful empirical predictive models. In this paper, we present significant enhancements of our Spatiotemporal Relational Probability Trees that enable autonomous discovery of spatiotemporal relationships as well as learning with arbitrary shapes. We focus our evaluation on two real-world case studies using our technique: predicting tornadoes in Oklahoma and predicting aircraft turbulence in the United States. We also discuss how to evaluate success for a machine learning algorithm in the severe weather domain, which will enable new methods such as ours to transfer from research to operations, provide a set of lessons learned for embedded machine learning applications, and discuss how to field our technique.

We introduce spatiotemporal relational probability trees (SRPTs), probability estimation trees for relational data that can vary in both space and time. The SRPT algorithm addresses the exponential increase in search complexity through sampling. We validate the SRPT using a simulated data set and we empirically demonstrate the SRPT algorithm on two real-world data sets.