Fast situation-based correction of AI systems

In this paper we present algorithms for continuous maintenance and improvement in a broad class Artificial Intelligence (AI) system whose primary function is to detect and report various multiple co-occurring objects or events in each data frame. The data frame combines features of various objects of interests as well as their relevant environments and as such represents a “situation”. Examples of such data frames and “situations” are feature vectors of deep neural networks with YOLO backbone. A distinct property of these data is that important operational information, such as input data corresponding to errors and misclassifications, does not always have fixed features associated with it. Instead, and depending on the context, this information can move from one feature to the other making the task of learning errors difficult. Here we present a solution to this problem by exploring clustered structure of data in high-dimensional spaces and analyse the effectiveness of these algorithms when presented with training samples of full input spaces. In addition to correcting errors, we test the outlined algorithms in the task of detecting adversarial attacks in a full input space. To illustrate the concepts in use, a case study is given which demonstrates the adaptive removal of false positives in an object-detection AI census system being developed for use in industry.