Towards Quantification of Assurance for Learning-Enabled Components

Quantifying Assurance in Learning-enabled Systems

“…Previously [7], we have described how assurance of ML components in an LES can be characterized through UQ of component-level properties associated with the corresponding (component-level) dependability attributes. Here, we extend our methodology to the system-level, relying on the following concepts: assurance is the provision of (justified) confidence that an item-i.e., a (learning-enabled) component, system, or service-possesses the relevant assurance properties.…”

“…1 not indicate whether or not it includes a mechanism to detect outliers or covariate shift. However, we include this variable here, motivated by our earlier work on componentlevel assurance quantification of the CNN [7], which revealed its susceptibility to outlier images. In fact, D models a runtime monitor for detecting out of distribution (OOD) inputs to the CNN.…”

“…To determine the emission probability Pr(CTE (t) a | I (t) ), first we used the Gaussian process (GP) model underpinning our prior work on component-level assurance quantification [7]. In brief, the idea is to use a GP to model the error performance of the CNN (i.e., its accuracy) on its input (i.e., runway images).…”

“…4 shows the assurance measure, Pr(|CTE (t=0) a | < offset) computed for two different offset values: 2m and 1.43m. 7 The interval [−2, 2] is a Bayesian credible interval within which the true CTE lies with probability ≈ 95%, based on Fig. 4.…”

“…When we use the DBN for runtime assurance, we implement it as a software component integrated into the LES. This can be viewed as an item to which we can apply our own assurance methodology, i.e., as in Section 2, and [7]. Thus, although we have not formulated assurance properties for the DBN, sensitivity and specificity are probabilistic performance metrics (albeit in a frequentist sense) that we can view as assurance measures in their own right, that we have now applied to our model.…”