An Evaluation Method for Generalization Errors of CNN using Training Data

Hyeon Ho Lee, Heung Seok Chae

http://doi.org/10.5626/JOK.2021.48.3.284

Even with high-performance CNNs, generalization errors, which are the errors on test datasets that are expected in the real world, are often high. This generalization error must be reduced so that the model can maintain its learned performance in the real world. This paper defines a response set as a neuron set that is frequently activated for each model class learned from the training dataset with high data diversity. Also, the differences in generalization errors due to the data diversity of the test dataset are considered. The difference is defined as a relative generalization error. In the current work, an evaluation method for CNN generalization error using only the training dataset is proposed by using the relationship between the CNN class response set and the relative generalization error. The case study confirms that the response set ratio is related to the relative generalization error and demonstrates the effectiveness of the evaluation method for generalization errors of CNN using training data.