Adaptive scheduling scheme for FPN-based explainable deep learning model

Abstract

Recently, AI technology provides information such as classification, recommendation, and prediction tousers and is being used for various purposes in a wide range of fields. However, AI technology has alimitation in that it does not provide validity for the final results derived due to the existence of blackbox characteristics. To solve this problem, many studies on the XAI (eXplainable AI) model are beingconducted to explain the transparency and validity of the process and judgment grounds judged by thedeep learning model. This XAI model improves reliability by providing users with explanation of whythese results were derived. However, in the case of the XAI model, there is a problem that resourcesare consumed as the number of objects to be processed increases. For example, when detecting objectsof various scales using a representative Feature Pyramid Network structure applied to improve explainability, visualized characteristics for each layer must be extracted, which consumes a lot of computingresources. To solve this problem, task scheduling shall be performed in consideration of the processingprocess in a heterogeneous environment with energy efficiency characteristics.To solve the above problem, this thesis considers the explanatory of the multi-scale feature map ofFPN and the scheduling considering the resource cost of visual explanatory map generation. First, weproposed explanatory modeling based on the visualization map of the Scalable Feature Block and theobject detection believer of the learned model. Since the explanatory properties of FPN Blocks differdepending on the characteristics of the target object, we modeled the explanatory properties as a linearmodel for the block reflection ratio of FPNs and detection performance by object. In addition, we proposean adaptive XAI task scheduling algorithm to minimize energy consumption of heterogeneous FPGAGPU resources when generating a Saliency Map. The proposed scheduling considers the distribution oftasks in sub-work units to minimize processing costs while satisfying the explanability. We evaluatedthe proposed algorithm through multiple object detection applications in heterogeneous FPGA-GPUclusters. Compared with existing accelerator-based scheduling, the proposed algorithm showed that itcould save 48% of energy consumption at the expense of less than 20% of processing time.