Research on Incremental Learning Methods Based on Sample and Category Prototype Playback
Jiamin Zhi, Yong Liu · International Journal of Computer Science and Information Technology · 2024
Machine learning technology has been successfully applied to computer vision, natural language processing and other fields, most of the existing machine learning models are deployed in actual production, because their categories and parameters are fixed, so they can only be applied to the categories that appear in the training set, and cannot incrementally learn the new categories that appear in the practical application. However, the advent of incremental learning has solved this problem. Therefore, this paper proposes an incremental learning method based on sample and category prototype replay, which aims to solve the forgetting problem in incremental learning while maintaining high accuracy and low computational complexity. Our method consists of two steps: sample playback and category prototype playback. In the sample playback stage, we record and store the historical data of the recent period, and in order to further reduce the category imbalance problem during playback, we propose a SMOTE algorithm based on global sample rationality, so that the model can learn the latest trends and changes. In the category prototype playback stage, we consider the importance of the representativeness of different samples and distinguish the important factors of different samples according to their dataset abundance, which makes it more accurate to calculate the prototype. The experimental results show that our method has good performance in both forgetting and prediction, and also has low computational complexity.