[3J1-GS-1] Fundamental AI, theory

[3J1-GS-1-05] Conditional selective inference for DNN and its applications

〇Daiki Miwa¹, Duy Nguyen Le Vo², Tomohiro Shiraishi³, Ichiro Takeuchi^3,2 (1. Nagoya Institute of Techonology, 2. RIKEN AIP, 3. Nagoya University )

Keywords:Selective Inference, Statistical Hypothesis testing, Saliency Map, DNN

Recently, Deep Neural Networks(DNNs) have been used widely in various fields. However, it is very important to consider the reliability of DNN models in problems where errors in decision making by DNN models are an important risk. We consider quantifying the reliability of prediction in the form of a p-values by statistical hypothesis testing. Unfortunately, conventional methods cannot provide valid p-values due to selection bias caused by the DNN model. In this study, we employ conditional selective inference framework, which has been actively studied recently. The proposed method is valid in the sense that it can control the probability of false positive detections. We proposed a novel algorithm and implementation with python, which can apply to various CNNs and enables us to conduct conditional selective inference without additional implementation cost. In addition, we report applying the proposed method to evaluate the reliability of the saliency maps by DNNs as examples.