在深度學習的領域中，分類任務的技術越趨成熟，而近年來相關研 究人員也陸續投身於具備新穎性檢測的階層式分類方法。我們在此篇 論文提出了一個利用分解信心值和連接條件機率，達到具新穎性檢測 的階層式分類模型，且訓練過程中不需加入額外的新類別資料，而基 準模型包含了自頂向下方法及攤平方法。將我們提出的模型與基準模 型相互比較，從結果可以得知，我們的模型除了有效提升已知類別的 準確度外，於尋找新的分類上也更加精準。此外針對階層式新類偵測 的任務，論文中提出了一個新的算分方法，目的是同時考慮新類偵測 以及階層式分類兩個任務，使其能更精確地顯示出模型的效能。;With the development of classification methods based on deep learning, hierarchical classification tasks with new class detection began to attract re searchers’ attention. In this paper, we propose a hierarchical classification with a novelty detection model by decomposing confidence and concatenating conditional probability, which can be trained without labeled novelty data. We compare it with a baseline model that combines the top?down method and flatten method. From the results, we found that our model can improve the classification accuracy of known categories and find instances belonging to new categories more effectively. We propose a new evaluation metric for the hierarchical novelty detection task. It considers both novelty detection and hierarchical classification so that it is able to express the performance of the model more obviously.
