A Deployment Method to Improve the Generalizability of Recurrent Neural Network
Yi Hu, Shuai Han, Shiji Wang, Cheng Li · 2024
The widespread adoption of deep learning models has inspired an urgent need for their generalization capabilities. Despite their impressive performance on training data, achieving high accuracy on deployed untouched data remains a daunting challenge. To address this problem, improving the model’s adaptability to new samples is imperative. In this paper, we delve into the metrics of deep learning models, pointing out that the upper bound of the generalization error is a quantitative measure of their generalization ability. Outlining methods to enhance this ability, we subsequently improve the LSTM model for modulation recognition using the identified upper bound on the generalization error and the outlined enhancement strategy, significantly improving the accuracy. Finally, an outlook on future research is provided.