[3Xin4] Poster session 1

[3Xin4-55] Method to Analyze the Factors of Uncertainty Caused by Insufficiency of Training Data

〇Yuxin Liang¹, Masashi Egi¹ (1.Hitachi, Ltd)

Keywords:Uncertainty Quantification, eXplainable AI, Data Centric AI, Uncertainty Attribution , MC Dropout

The quality of training dataset is important for improving the accuracy of AI. To improve the quality, we are investigating a method to identify data requirements on augment data. Uncertainty Quantification technology is applicable for the purpose to estimate uncertainty of prediction caused by the scarcity on training data (Epistemic uncertainty). To identify data requirement on augmented data, the technology to analyze the factors which increase or decrease the Epistemic uncertainty is demanded. Conventional method is specialized for recognition problem. We apply the conventional method to analyze factors that increase or decrease Epistemic uncertainty for regression problem. We experiment the effectiveness of proposed method on the train dataset which omit train data with specific feature conditions for intentioned Epistemic uncertainty factors. The experiment result shows that the proposed method is effective. We get prospect to realize the quality improvement of training data.