详情

原文传递 Leveraging Subject Matter Expertise to Optimize Machine Learning Techniques for Air and Space Applications.

题名：	Leveraging Subject Matter Expertise to Optimize Machine Learning Techniques for Air and Space Applications.
作者：	Cho, P. Y.
摘要：	In this research, we develop machine learning and statistical methods that are tailoredfor Air Force applications through the incorporation of subject matter expertise.In particular, we develop techniques for incorporating subject matter knowledge inneural networks, Bayesian regression, and structural causal models. These techniquesare developed in the context of three separate application areas: localizing pointdefects in transmission electron microscopy (TEM) of crystalline materials; estimatingthe relationship between attributes of fighter pilot communities and flight mishap rate;and analyzing Air Force evaluation process.Our first contribution is a novel method for localizing point defects in TEM im-ages of crystalline materials using principal component analysis (PCA) and a convolu-tional neural network (CNN). Notably, the design of the PCA-CNN method leveragesknowledge about point defects in crystalline materials. Furthermore, the method isa self-supervised method that is trained without labeled images of point defects and,thus, represents a novel methodological contribution. We show that the tailored PCA-CNN method outperforms CutPaste, a state-of-art artificial intelligence (AI) modelfor defect localization method, on both simulated and experimental TEM images.Our second contribution reveals the relationship between attributes of fighter pilotcommunities and flight mishap rates through the use of predictive projection andBayesian regression. We use personnel and mishap data from 2007-2020 to present anin-depth analysis of historic trends within fighter pilot communities. In our analysis ofhistoric mishap data, we find evidence of abnormal mishap cost estimation behaviornear the threshold between class B and C mishaps. Using Bayesian regression withfeature selection via predictive projection, we find that pilot communities with higheraverage flight hours in the last year are associated with reduced mishap rates. Ahigher percentages of pilots who are DGs, are IPs, and have advanced academicdegrees are also associated with reduced mishap rates.Lastly, we demonstrate the use of Bayesian priors to incorporate the subject mat-ter knowledge gained from prior qualitative studies on mishap safety. Our thirdcontribution provides a framework for estimating causal effects using data associatedwith Air Force evaluation processes. Air Force evaluation processes are unique be-cause the causal relationships that induce the observed data are often known. Forexample, policy dictates which factors can and cannot be considered by a promotionboard. We use structural causal models to represent our knowledge of the evalua-tion processes. Under the assumption of a linear causal model, we derive a formula,ˆβ = C −1x⃗C ′xy , for computing the coefficients in a regression via the pair-wise covari-ances of the predictors. This allows for the estimation of causal quantities pertainingto evaluation processes via regression and the proper selection of controls.
总页数：	129 pages