부산대학교 도서관

The scope of AI application is expanding to mission-critical domains such as safety, healthcare and finance. To introduce AI in such domains, understandability of the prediction logic is required. However typical accurate models such as deep neural networks are too complicated to understand. Although there exist methods for generating globally understandable models, their accuracy and applicability are often insufficient. To achieve high accuracy by understandable models for classification and regression tasks, we propose piecewise simple function (PSF) as a general understandable model and a method for training accurate PSF. PSF consists of a few local submodels and can be regarded as a generalization of the piecewise affine model. However, the output of the proposed method is expected to be more accurate since it utilizes information of accurate black-box AI. More concretely, we specify the partition of regions on the basis of information of black-box AI such as important features and thresholds. The proposed method is based on the hypothesis that black-box AI is often overly complicated in some regions and can be replaced with simpler models. Through numerical experiments, we demonstrate that PSF often achieves almost the same accuracy as original black-box AI.