Memory Using Data Generator in Continual Learning for Remote Sensing Scene Classification
Nassim Ammour · 2021
Deep learning models suffer from catastrophic forgetting and collapse dramatically when they are subjected to continual learning process. To overcome this handicap, we propose a novel continual learning approach based on previously seen data auto-generation sub-networks. The proposed model continually learns a set of sequential classification tasks or classes, where each classification task includes a certain number of remote sensing scenes or classes. The proposed neural networks architecture is composed of two trainable sub-networks. The first module adjusts its weights by minimizing the discrimination between the land-cover classes error during the new task learning. In parallel, the second module attempts to learn how to reproduce the task data by discovering the latent data structure of the new task dataset. Experiments are conducted on Merced dataset. The experimental results confirm the outperformance and robustness of the proposed model.