¿µ¹®³»¿ë (English Abstract) |
Recent incremental learning studies have revealed that adding prompts to input data is effective in incremental learning settings, utilizing prompts that have been actively used in natural language processing. However, when analyzed how prompts were used for learning, there is a problem that existing method do not learn prompts efficiently. Therefore, in this paper, we propose a method which modifies the criteria to select prompts which are added to an input sample. Also, we introduce Assistant prompts to training process by utilizing an Assistant layer, so that the prompts can be used more efficiently for incremental learning. Through the experiment, we show that our model which has higher performance while using less parameters than before, can solve previous problem. |