Introducing Class Replacement Technique in Class Incremental Learning in Generative Models
Taro Togo, Ren Togo, Keisuke Maeda, Takahiro Ogawa, Miki Haseyama · 2024
This paper presents a novel method for Class Incremental Learning (CIL) in image-generative models, using a class replacement technique. There has been little focus on unnecessary classes in CIL for image-generative models. Our approach is inspired by human forgetting and cognitive processes. By replacing unnecessary classes with noise images, we improve the performance of learning new classes in CIL. We performed fine-tuning with replacing techniques in MNIST and investigated the impact of replacing unnecessary classes. Through this study, we aim to promote the continual use of image-generative models by replacing unnecessary classes.