Join Pacific Northwest National Laboratory at the Joint Statistical Meetings in Boston, Massachusetts!
Graphic created by Kelly Machart | Pacific Northwest National Laboratory
Boston, Massachusetts
One of the largest statistical events in the world, the Joint Statistical Meetings (JSM) will be held in Boston, Massachusetts, and will feature 600+ sessions in the areas of statistics, data science, and enriching society through AI. The event anticipates more than 5,000 attendees from around the world, including several researchers and mathematicians from Pacific Northwest National Laboratory (PNNL).
Uncertainty Quantification in Multimodal AI with Thoughts on Human Interpretability
PNNL Author: Karl Pazdernik
Coauthors: Brendan Kennedy and Jessica Baweja
Abstract
In the critical task of making generative models trustworthy and robust, methods for uncertainty quantification (UQ) have begun to show encouraging potential. However, many rely on rigid heuristics that fail to generalize across tasks and modalities. We introduce Directional Concentration Uncertainty (DCU), a novel statistical procedure based on the von Mises–Fisher distribution that uses continuous embeddings to measure the geometric dispersion of multiple generated outputs. DCU captures uncertainty without task-specific heuristics and achieves calibration performance comparable to or exceeding that of prior methods like semantic entropy, while generalizing well to complex multimodal tasks. We present a framework for DCU’s integration into broader UQ strategies for multimodal and agentic systems. Finally, UQ output must be interpretable to be trustworthy so that the way uncertainty is described and visualized can be as critical as the metric itself. To that end, we include findings from 90-minute virtual interviews conducted with nonproliferation analysts to explore how uncertainty should be communicated to best support decision-making in high-consequence environments.
A Hierarchical Multistate Model for Juvenile Salmonid Survival and Migration in the Columbia River
PNNL Author: Narmadha Mohankumar
Coauthors: Ryan Harnish, Katherine Deters, Tao Fu, Scott Titzler, Jill Janak, Jayson Martinez, Robert Mueller, Adam Hall, and Zhiqun Deng
Abstract
The survival and migration dynamics of juvenile salmonids in hydroelectric systems are governed by a complex interplay of environmental conditions and dam operations. Statistical inference in this setting is complicated by imperfect detection and uncertainty in individual passage times. We address these challenges using a time-integrated migration and survival (TIMS) model to analyze yearling Chinook salmon (Oncorhynchus tshawytscha) and steelhead (O. mykiss) behavior and survival at Little Goose Dam (LGS) in the Columbia River hydro system. TIMS is a hierarchical multistate mark-recapture framework that embeds a migration submodel for travel time between detection arrays, enabling survival and migration rates to be modeled as explicit functions of time-varying covariates, accounting for unobserved passage times. Using acoustic telemetry data from 2012, 2018, and 2025, we evaluate the effects of route, fish length, water velocity, spill, and temperature across reaches from upstream of LGS to McNary Dam. This framework yields robust estimates of survival and migration under dynamic conditions, improving inference of operational impacts on biological performance in regulated rivers.
As a national laboratory that conducts an abundance of research using advanced mathematics, we are always searching for talented individuals looking to be a part of our mission.