[4I2-J-2] Machine learning: uncertainty of targets

[4I2-J-2-01] Importance of Uncertainty Estimation in Deep Learning

〇Iwao Maeda¹, Hiroyasu Matsushima¹, Hiroki Sakaji¹, Kiyoshi Izumi¹, David deGraw², Hirokazu Tomioka², Atsuo Kato³, Michiharu Kitano³ (1. The University of Tokyo, 2. Daiwa Securities Co. Ltd., 3. Daiwa Institute of Research Ltd.)

Keywords:Deep Learning, Uncertainty, Image Recognition

In recent years, predictions by machine learning and deep learning methods are utilized in various scenes of society. A model trained with deep learning methods can predict the target with high accuracy, but can not consider the predictive confidence sufficiently, and may predict high confident for extrapolated data which is hard to predict. In this study, we applied ordinary deep learning methods and methods considering predictive uncertainty, proposed in recent years, to an image classification task, and verified the robustness of trained models against extrapolated data. Models trained with the ordinary deep learning methods predicted high confidence values for data having characteristics not existing in the training data, but models trained with the methods considering uncertainty predicted low confidence values for such data. By using methods considering uncertainty, it is possible to avoid mispredictions for extrapolated data. Experimental results suggest the importance of uncertainty estimation in deep learning.