AI Maintenance: A Robustness Perspective

In general, the performance of an AI model is Just like the indispensable role of cars in the evaluated in the average case, by comparing the modern world, AI-empowered technology, and model predictions on a set of data samples to their ML-based systems and algorithms are bringing ground-truth labels and then using the average revolutionary changes and far-reaching impacts prediction result as a performance metric, such on our life, society, and environment, if not as the top-1 classification accuracy measuring the happening already. As AI models are perceived fraction of correct model prediction on the mostlikely as a new "vehicle" to a better future, this article (top-1) class over a dataset. In contrast, the aims to stress the importance of formalizing and adversarial scenario evaluates the model performance practicing AI maintenance from the robustness in the worst case among all possible and perspective, by drawing analogies in the model plausible changes (often pre-specified) to the data development and deployment between car and AI. and AI model, by assuming a virtual adversary is Towards achieving trustworthiness and sustainability in place. Moreover, the unseen scenario evaluates for AI, this article is motivated by the following the model performance on new data samples that question: Cars require regular inspection, are drawn from a different data distribution than maintenance, and continuous status monitoring, the seen data samples during training (but not why should AI technology be any different?