A Fast Training Method using Bounded Continual Learning in Image Classification
Seunghui Jang, Yanggon Kim · 2021
These days, Deep neural networks (DNNs) are showing good performance in image classification. They bring sufficient performance not only in specific fields such as medical images and meteorological observation images, but also in fields necessary for daily life. However, for them to be more useful practically, they should be able to add new tasks to suit a changing environment. This is the same as saying that they should be able to learn by adding new data. Unfortunately, since catastrophic forgetting is an inevitable feature of connectionist model, it is very difficult to capture both accuracy and computational efficiency in continuous task learning. In this paper, we propose Bounded Continual Learning (BCL) based on inductive transfer learning. BCL extracts feature values from sub-models created by separating base datasets and train a new classifier. BCL showed very good performance in training time efficiency to learn tasks in a sequential. We demonstrate our approach is flexible and efficient by various classification tasks based on the CIFAR dataset.