Exploration of continual learning based on the comparison and analysis of LWF and EWC
Zihang Wang · Applied and Computational Engineering · 2023
Nowadays, continual learning is seen as one of the important fields of strong artificial intelligence. People focus on it and propose a lot of methods in order to alleviate catastrophic forgetting problem which is serious and essential and cannot be avoided. Among all the methods, EWC and LWF are two typical methods and they have some similarities and differences. This paper adopts comparison method to research the effect of LWF and EWC applied on different datasets and also compares some similar methods like R-EWC and LFL. Meanwhile, some analysis and suggestions about these methods are also given. From the results, both methods improve memory ability. But LWF performs better in testing old knowledge. In contrast, EWC has a stronger ability to learn new knowledge. Both EWC and LWF have some limitations. There are some problems which cause errors for the principle of EWC while for LWF, the confirmation of hyper-parameters has great influence on the experimental results and the method itself has some errors.