Sandia Labs FY22 Laboratory Directed Research & Development Annual Report

FY22 ANNUAL REPORT

IMPROVING ACTIVE LEARNING FOR LANGUAGE MODEL TO COMBAT DISINFORMATION.

Various supervised language-modeling techniques can help decipher and dismantle disinformation campaigns, but they rely heavily on large, human labeled training datasets. This LDRD addresses this problem by deploying active learning (AL) to generate labeled datasets and optimize human input. Task adaptive pretraining was used to leverage the unlabeled data and boost the performance of the classifier. A disinformation

was developed to save human time from being wasted on iterations that would not significantly benefit classifier performance. This work laid the foundation to continue using AL in the disinformation data annotation domain and to shift the AL paradigm to optimize human input. This project was sponsored by Sandia Fellow Bill Miller. (PI: Emily Kemp)

rhetoric metric was created to measure the presence of common disinformation rhetoric, for both the classifier and human to leverage. The DRM was combined with an uncertainty criterion to create a hybrid acquisition method. A sophisticated and robust stopping strategy

Averaged AL simulation results.

EXPERIMENTAL QUANTUM-ENABLED, SUPER-RESOLUTION IMAGING TECHNIQUE IS A “

Rayleigh diffraction limits the obtainable imaging resolution from conventional imaging techniques by making it hard to discriminate an object’s geometrical features smaller than the wavelength of probe light. Rayleigh-diffraction-limited imaging resolution is explained in quantum mechanics as a consequence of light-noise fluctuation (i.e., shot noise). Sandia developed a super-resolution quantum imaging technique where a much smaller (enhanced) imaging resolution can be obtained by employing squeezed light, where light noise is significantly reduced below the shot noise. This

LDRD team advanced the technology by delivering squeezed light with 10 dB less noise than the shot noise. Such a large noise reduction allows a significant advance in performance as compared to Rayleigh-diffraction-limited conventional imaging. This quantum imaging technique will be a game changer in biology and chemistry, where super resolution optical diagnostics will elucidate the underlying physical processes. This project was co-sponsored by two Sandia Fellows – Ed Cole and Dave Chandler. (PI: Daniel Soh)

80