From Forensics to GANs: An Early-Career Statistician’s Experience Doing Research in a Government Lab (306480)Ruthmara Corzo, National Institute of Standards and Technology
Michael Frey, National Institute of Standards and Technology
Felix Jimenez, University of Colorado, NIST
*Amanda Koepke, National Institute of Standards and Technology
Eric Steel, National Institute of Standards and Technology
Keywords: forensic science, GAN, machine learning, early-career, statistical collaboration
Working in a government lab provides exposure to many exciting research opportunities. As a statistician at NIST, I apply my skills to a variety of interesting applications, two of which I detail here. First, in the forensic analysis of glass fragments, a new fragment is compared to samples from a known glass source to determine whether they originated from the same source. Current approaches do not address the effect of fragment shape on the elemental fluorescence intensity measured using micro-XRF spectrometry. We develop a framework to identify matching samples from the spectra of fragments and study the sensitivity of the results to differences in fragment shape. Second, artificially generated data that accurately reflect the underlying distribution of real data are essential to many data analytic applications. Recently developed generative adversarial networks allow for creation of an unlimited number of observations from an unknown distribution given a finite sample. The data generator’s success depends on the size of the available sample. Focusing on low-dimensional data generation, I present bounds on this performance and its variability as a function of sample size.