34 Association of Research Libraries Research Library Issues 299 2019 for designers to adjust their models (that is, it is an error-correction process), allowing users to examine feature selection explanations would provide a justification for why a decision was made and would allow them a basis to challenge that result. In seeking explanations, people rarely ask for or rely on complete explanations. Rather than reviewing and assessing all the causes (even if provided), people tend to be highly selective. We seek and accept explanations that “satisfice.” Approximation or abstraction are techniques that create a more simplified model to explain the more complex model. Approaches such as model distillation or model-agnostic feature reduction create a simplified presentation of the algorithmic model. This approximation or abstraction may compromise accuracy, but it provides an accessible representation that enhances understandability. XAI researcher Trevor Darrell believes that “the solution to explainable AI is more AI.”14 In this approach to explanation, oversight AI are positioned as intermediaries between an AI and its users. These AI have been called “ethical governors,”15 “flight data recorders,”16 and, more ominously, “AI guardians.”17 These examples of intelligent middleware offer the ability to interpolate the values and expectations of third parties, such as research libraries, in the process of deriving an explanation from an AI. Replication is a recognized verification strategy in many aspects of research. Being able to independently reproduce results in different settings provides evidence of veracity and supports user trust. However, documented problems in reproducing machine-learning research have questioned the generalizability of these approaches and undermined their explanatory capacity. In response, a “Reproducibility Challenge” was created by the International Conference on Learning Representations (ICLR) to validate 2018 and 2019 conference submissions.18 More rigorous replication through the availability of all necessary components will be important to this type of verification.
Previous Page Next Page