[4Xin1] Poster session 2

[4Xin1-80] Predicting classification accuracy in image classification using CNN

〇Mizuki Dai¹, Knya Jin'no¹ (1.Tokyo City University)

Keywords:CNN, Classification accuracy prediction

To examine the extent to which the test classification accuracy of a CNN after training can be predicted by the initial training results.The design of CNN models is often built using empirical rules. This requires a very large number of experiments to build the model. If the test classification accuracy after training can be predicted by the accuracy in the early stages of training, the time required to build the model can be reduced. In this study, the explanatory variables were the values we focused on, such as the value of classification accuracy in the early stages of learning, and the objective variable was the classification accuracy after learning, which was predicted using multiple regression analysis. As a result of evaluating the predicted values using RMSE, it was confirmed that the classification accuracy after learning can be predicted with very high accuracy. The results of this study are expected to be used as an evaluation indicator for NAS.