Publications

In several mission contexts, it is desirable to estimate the performance of large language models (LLMs) on tasks that we cannot run directly. In light of published “scaling laws” our hypothesis is that some tasks should be consistently more challenging than others based on characteristics of the task. The goal of this project was to begin quantifying how much information about LLM performance can be gained from the features of a model and a task. Two of our statistical models struggled to converge. Pass/fail test results may provide limited information for inference beyond model quality and task difficulty, but we see no evidence at this time for significant feature interaction effect sizes, arguing for simple models. Future work extending the models to capitalize on perplexity of ground truth answers is suggested. This project also introduces “Depth of Knowledge Variant Testing” as a strategy for more finely assessing language models on open domain question and answer tasks. We developed sets of questions that ask a language model to produce similar information while demonstrating increasing depth of knowledge, and also relabeled existing Q&A test questions with their depth of knowledge. Our results suggest further consideration of Bloom’s taxonomy and further refinement of prompts to properly elicit information at varying depths. In the course of this work, we set up a basic infrastructure for standardizing tasks and testing many language models on these tasks. In addition to testing the predictive quality of model features and performance across test suites, with this project we have introduced two new task features to contextualize each test question: the Dewey Classification main category of information covered, and the Bloom’s taxonomy level that corresponds to the depth of knowledge probed by the question. Splits across these and other features produced over five hundred task subtypes with distinct feature vectors, which we tested on half a dozen models.

High-quality uncertainty quantification (UQ) is a critical component of enabling trust in deep learning (DL) models and is especially important if DL models are to be deployed in high-consequence applications. Conformal prediction (CP) methods represent an emerging nonparametric approach for producing UQ that is easily interpretable and, under weak assumptions, provides a guarantee regarding UQ quality. This report describes the research outputs of an Exploratory Express Laboratory Directed Research and Development (LDRD) project at Sandia National Laboratories. This project focused on how best to implement CP methods for DL models. This report introduces new methodology for obtaining high-quality UQ from DL models using CP methods, describes a novel system of assessing UQ quality, and provides experimental results that demonstrate the quality of the new methodology and utility of the UQ quality assessment system. Avenues for future research and discussion of potential impacts at Sandia and in the wider research community are also given.

Abstract not provided.

Accurate understanding of the behavior of commercial-off-the-shelf electrical devices is important in many applications. This paper discusses methods for the principled statistical analysis of electrical device data. We present several recent successful efforts and describe two current areas of research that we anticipate will produce widely applicable methods. Because much electrical device data is naturally treated as functional, and because such data introduces some complications in analysis, we focus on methods for functional data analysis.

Deep learning (DL) models have enjoyed increased attention in recent years because of their powerful predictive capabilities. While many successes have been achieved, standard deep learning methods suffer from a lack of uncertainty quantification (UQ). While the development of methods for producing UQ from DL models is an active area of current research, little attention has been given to the quality of the UQ produced by such methods. In order to deploy DL models to high-consequence applications, high-quality UQ is necessary. This report details the research and development conducted as part of a Laboratory Directed Research and Development (LDRD) project at Sandia National Laboratories. The focus of this project is to develop a framework of methods and metrics for the principled assessment of UQ quality in DL models. This report presents an overview of UQ quality assessment in traditional statistical modeling and describes why this approach is difficult to apply in DL contexts. An assessment on relatively simple simulated data is presented to demonstrate that UQ quality can differ greatly between DL models trained on the same data. A method for simulating image data that can then be used for UQ quality assessment is described. A general method for simulating realistic data for the purpose of assessing a model’s UQ quality is also presented. A Bayesian uncertainty framework for understanding uncertainty and existing metrics is described. Research that came out of collaborations with two university partners are discussed along with a software toolkit that is currently being developed to implement the UQ quality assessment framework as well as serve as a general guide to incorporating UQ into DL applications.